From 5fe457ad0179c1f6f68e041a068587254245e17e Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 14:49:21 -0700 Subject: [PATCH 01/16] Studio: bound how many tool approvals may park their slot (#7496) * Bound how many approvals may park, against the executor #7455 landed parking, which is the right shape and supersedes what this branch was carrying. It is unbounded, though, and the thing it is unbounded against is not the GPU. A run stopped on an approval prompt is blocked inside the to_thread(next, gen) call that drives it, so it holds one of asyncio's default min(32, cpu + 4) executor threads until the user answers. The slot cap used to bound that. Parking hands the slot back, which admits another run that can park too, so the ceiling became the wait line: 64 deep on a 1-slot backend. Long before that, the executor is full and nothing else in the backend runs, including generation steps for chats that already hold slots and the stream teardown that would clean up after a disconnect. The pool already permits `capacity` pending prompts, and each park adds one more, so the budget is what the executor has left after the cap and a reserve of 4. On this machine (32 workers) --parallel 4 gets 8 parks and 20 free threads, --parallel 24 gets 4 and 4, and --parallel 28 or higher gets none: there the prompt keeps its slot and behaves exactly as it did before parking existed. Counted process-wide rather than per queue. There is one executor, but a per-queue budget is the same allowance again for every backend, and base_url carries a fresh port on every model load, so a reload would mint a queue that knows nothing about the approvals still parked on the old one. A reset clears it too, or a leaked claim shrinks the budget for the life of the process. park() reports whether it took the budget, and a refusal costs nothing to undo because the slot never left its holder. The stream reads that answer rather than recording a refused park as parked, which would make it skip the park for every later approval in the same run even once the budget freed up. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Size the park budget from the executor's own CPU count Two review findings, both real. The budget read os.cpu_count(). 3.13 sizes ThreadPoolExecutor from os.process_cpu_count(), which honours CPU affinity and cgroup quotas, and asyncio's default executor is a plain ThreadPoolExecutor(), so a container pinned to one core on a 64-core host got a 5-thread executor and a budget computed from 64. The bound was then looser than no bound at all in exactly the environment that can least afford it. It asks the same source the executor does, and the test compares against a real ThreadPoolExecutor rather than restating the formula, so it stays right on 3.12 as well. The reserve was a flat 4, which on that same 5-thread executor left nothing to budget and turned parking off entirely. Small hosts are where a chat most needs to keep moving while another sits on a prompt. It scales now, and the ceiling has a floor of two: a quarter of five is one, and one park cannot cover two chats on prompts at once, which is what #7455's own two-approvals test needs. Without that floor, that test fails on a one or two CPU runner. `spare` still takes the budget to zero when the pool already fills the executor, so nothing about a 32-worker machine changes: --parallel 4 still gets 8 parks, 24 gets 4, 28 gets none. The two behavioural budget tests pin the worker count rather than reading it off the runner, and the property test sweeps executor sizes from one CPU to 64 instead of asserting against whatever the host happens to have. The whole suite passes with the CPU count faked to 1, 2 and 4, which is how both of these were reproduced. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Size the park budget from every live backend, and free it on the answer Two review findings, both real. The budget was global but sized from one queue's capacity. A reload mints a queue on a new port while the old one drains, so both are live, and prompts on both park executor threads. Eight parks on an old 1-slot queue plus a new 24-slot backend is 32 threads on a 32-thread executor, with the new backend's prompts refused and holding their slots, which is the state the reserve exists to prevent. It sums the capacity of every backend still serving instead. Idle queues are skipped: those are the ones the registry is about to evict, and they are holding nothing. The budget also outlived the wait it was paying for. unpark_async only dropped it after reacquiring a slot, but the generator yields its post-approval event first, so the executor thread is already back in the pool while the resume queues. An approved chat waiting on a slot would refuse a different chat's park, and that chat then keeps the slot the resumer is waiting for, so an unanswered prompt strands chats that were already approved. The budget is released when the prompt wait ends now, and the queue's parked count still runs until the slot is back, which is what guards idle eviction and the resume ordering. Both are separate counters on the lease as a result, and every exit from a park drops the budget: unpark, unpark_async and release. That last one was the mutant that came back missed, since a client disconnecting on a prompt releases straight out of parked and would otherwise lose a budget slot for good. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Tighten the comments on the park budget --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --- .../backend/core/inference/llama_admission.py | 133 ++++++++++- studio/backend/routes/inference.py | 5 +- studio/backend/tests/test_llama_admission.py | 226 ++++++++++++++++++ 3 files changed, 355 insertions(+), 9 deletions(-) diff --git a/studio/backend/core/inference/llama_admission.py b/studio/backend/core/inference/llama_admission.py index db9a5d8ce4..7bf0dd7429 100644 --- a/studio/backend/core/inference/llama_admission.py +++ b/studio/backend/core/inference/llama_admission.py @@ -58,6 +58,80 @@ DEFAULT_ADMISSION_QUEUE_PER_SLOT = 16 DEFAULT_ADMISSION_MIN_QUEUE = 64 +def _executor_workers() -> int: + """Threads asyncio's default executor runs to_thread work on. + + Mirrors ThreadPoolExecutor's own default sizing, which is what + ``run_in_executor(None, ...)`` builds. 3.13 sizes it from + ``process_cpu_count()``, which honours CPU affinity and cgroup quotas; + ``cpu_count()`` would budget from the whole host inside a one-core container. + """ + cpus = getattr(os, "process_cpu_count", os.cpu_count)() or 1 + return min(32, cpus + 4) + + +def _executor_reserve(workers: int) -> int: + """Threads kept clear of parked approvals, for generation steps, stream + teardown and unrelated to_thread work. Scaled rather than flat: a flat count + would leave a 5-worker executor (one usable CPU) no budget at all. + """ + return max(2, workers // 8) + + +def _max_parked(capacity: int) -> int: + """How many holders may sit on an approval prompt with their slot given back. + + A pending prompt parks an executor thread (the loop blocks inside + to_thread(next, gen)) whether or not it parked its slot, the pool already + permits `capacity` of those, and every park admits one more, so budget only + what the executor has left over. Zero on a backend whose --parallel alone + fills it: the prompt then holds its slot, as it did before parking existed. + """ + workers = _executor_workers() + spare = workers - _executor_reserve(workers) - max(0, capacity) + # A quarter of the executor, floored at two while `spare` allows: a quarter of + # five is one, and one park cannot cover the two simultaneous prompts #7455 + # exists for. + return max(0, min(max(2, workers // 4), spare)) + + +# Process-wide, not per queue: there is one executor, and base_url takes a fresh +# port on every load, so a per-queue budget would hand the same allowance to each +# backend and to every reload, blind to the approvals parked on the old queue. +_PARK_LOCK = threading.Lock() +_parked_total = 0 + + +def _claim_park(limit: int) -> bool: + global _parked_total + with _PARK_LOCK: + if _parked_total >= limit: + return False + _parked_total += 1 + return True + + +def _drop_park() -> None: + global _parked_total + with _PARK_LOCK: + _parked_total = max(0, _parked_total - 1) + + +def _live_capacity(current: "LlamaAdmissionQueue") -> int: + """Slots across every backend still serving requests. + + One queue's capacity is the wrong denominator for a budget sized against the + one executor: a reload drains the old queue alongside the new one, and + prompts on both park threads. Idle queues hold nothing and are about to be + evicted. + """ + with _QUEUES_LOCK: + queues = list(_QUEUES.values()) + # is_idle takes each queue's own lock, so never while holding _QUEUES_LOCK. + total = sum(queue._capacity for queue in queues if queue is current or not queue.is_idle()) + return total if any(queue is current for queue in queues) else total + current._capacity + + @dataclass(frozen = True, **_SLOTS) class LlamaAdmissionConfig: enabled: bool = DEFAULT_ADMISSION_ENABLED @@ -214,7 +288,7 @@ class _Waiter: class LlamaAdmissionLease: - __slots__ = ("_queue", "_slot", "_released", "_release_lock", "_parked") + __slots__ = ("_queue", "_slot", "_released", "_release_lock", "_parked", "_budgeted") def __init__( self, @@ -226,27 +300,52 @@ class LlamaAdmissionLease: self._released = False self._release_lock = threading.Lock() self._parked = False + self._budgeted = False @property def slot(self) -> Optional[int]: """Pool slot this lease holds, or None when admission is disabled.""" return self._slot - def park(self) -> None: + def park(self) -> bool: """Hand the slot back while this holder waits on something off the GPU. A run stopped on a tool approval prompt is not decoding, so holding its slot would let unanswered prompts fill the pool while llama-server idles. The lease itself stays valid: releasing it after a park is still correct. + + False when the park budget is spent and nothing was given back: the + caller keeps its slot across the prompt, as it did before parking + existed. Slower for whoever is behind it, but each freed slot admits + another run that can park too, on the executor the generators run on. """ queue = self._queue - slot = None with self._release_lock: if queue is None or self._released or self._parked: - return + return False + # Under the lease lock so the decision and the handover cannot split. + # Nothing takes the queue lock then a lease lock, so this order is + # the only one in play. + if not queue.try_park(self._slot): + return False self._parked = True - slot, self._slot = self._slot, None - queue.park(slot) + self._budgeted = True + self._slot = None + return True + + def _drop_budget(self) -> None: + """Give the executor budget back now the prompt wait is over. + + Separate from the queue's parked count, which lasts until the slot is + back: the executor thread is free the moment the answer arrives. Holding + the budget until the resume lands would refuse someone else's park for a + finished wait, and that someone holds the slot the resumer wants. + """ + with self._release_lock: + if not self._budgeted: + return + self._budgeted = False + _drop_park() def unpark(self) -> None: """Drop the parked state without reclaiming a slot. @@ -259,6 +358,7 @@ class LlamaAdmissionLease: if not self._parked: return self._parked = False + self._drop_budget() if self._queue is not None: self._queue.unpark() @@ -278,6 +378,9 @@ class LlamaAdmissionLease: queue = self._queue if queue is None or not self._parked: return + # Before the wait, not after: the prompt is answered, so this holder is + # already off the executor and must not keep anyone else off it. + self._drop_budget() slot = await queue.acquire_parked_slot(cancel_event = cancel_event, poll_s = poll_s) stranded = None with self._release_lock: @@ -304,6 +407,7 @@ class LlamaAdmissionLease: self._released = True queue = self._queue parked, self._parked = self._parked, False + self._drop_budget() if queue is not None: if parked: queue.unpark() @@ -513,12 +617,20 @@ class LlamaAdmissionQueue: self._release_slot_locked(slot) self._grant_waiters_locked() - def park(self, slot: Optional[int]) -> None: - """Return a parked holder's slot to the pool. See ``LlamaAdmissionLease.park``.""" + def try_park(self, slot: Optional[int]) -> bool: + """Return a parked holder's slot to the pool. See ``LlamaAdmissionLease.park``. + + False leaves the slot with its holder, so a refused park costs nothing to + undo. The per-queue count is only what ``is_idle`` reads; the budget and + the capacity it is sized from are both process-wide. + """ + if not _claim_park(_max_parked(_live_capacity(self))): + return False with self._lock: self._parked += 1 self._release_slot_locked(slot) self._grant_waiters_locked() + return True def unpark(self) -> None: with self._lock: @@ -684,5 +796,10 @@ def get_llama_admission_queue(key: str) -> LlamaAdmissionQueue: def reset_llama_admission_queues() -> None: + global _parked_total with _QUEUES_LOCK: _QUEUES.clear() + # The budget outlives the queues it was claimed against, so dropping them + # without it leaks the count and shrinks the budget for good. + with _PARK_LOCK: + _parked_total = 0 diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py index 8b15779a50..53b4136e32 100644 --- a/studio/backend/routes/inference.py +++ b/studio/backend/routes/inference.py @@ -9413,7 +9413,10 @@ async def openai_chat_completions( if lease is None: return if on: - lease.park() + # Refused when the budget is spent: the slot stays here, + # so there is nothing to take back afterwards. + if not lease.park(): + return elif wait: # Resuming: park() may have handed our slot to a waiter, so wait for room instead # of putting two holders on one slot. diff --git a/studio/backend/tests/test_llama_admission.py b/studio/backend/tests/test_llama_admission.py index 9ff19ec27d..1b1aeb1cc5 100644 --- a/studio/backend/tests/test_llama_admission.py +++ b/studio/backend/tests/test_llama_admission.py @@ -1066,3 +1066,229 @@ def test_an_immediate_arrival_cannot_take_an_approved_chats_slot(): assert queue.snapshot().active <= 1 asyncio.run(scenario()) + + +def test_parking_is_bounded_so_the_thread_pool_cannot_be_drained(monkeypatch): + # A pending prompt parks an executor thread (the loop blocks inside + # to_thread(next, gen)) and frees a slot that admits another run which can + # park too, so unbounded parking drains the pool the generators run on. + # Pinned because the real budget follows the runner's usable CPUs. + monkeypatch.setattr(llama_admission, "_executor_workers", lambda: 32) + + async def scenario(): + queue = get_llama_admission_queue("http://llama.test") + config = LlamaAdmissionConfig() + limit = llama_admission._max_parked(1) + assert limit >= 1 + + leases = [] + for _ in range(limit): + lease = queue.reserve(capacity = 1, config = config).lease_nowait() + assert lease is not None and lease.park() + leases.append(lease) + + refused = queue.reserve(capacity = 1, config = config).lease_nowait() + assert refused is not None + assert not refused.park(), "parking is unbounded" + # Refusing means keeping the slot, the old behaviour, not an error. + assert refused.slot is not None + assert queue.snapshot().active == 1 + + leases[0].unpark() + assert refused.park(), "budget was not returned" + for lease in leases[1:] + [refused]: + lease.release() + leases[0].release() + + asyncio.run(scenario()) + + +def test_the_park_budget_is_shared_by_every_queue(monkeypatch): + # One executor, so a per-queue budget would be handed out again to every + # backend and to every reload onto a fresh ephemeral port. + monkeypatch.setattr(llama_admission, "_executor_workers", lambda: 32) + + async def scenario(): + config = LlamaAdmissionConfig() + first = get_llama_admission_queue("http://llama.test:1") + second = get_llama_admission_queue("http://llama.test:2") + limit = llama_admission._max_parked(1) + + for index in range(limit): + queue = first if index % 2 == 0 else second + lease = queue.reserve(capacity = 1, config = config).lease_nowait() + assert lease.park() + + spare = second.reserve(capacity = 1, config = config).lease_nowait() + assert not spare.park(), "each queue got its own budget" + + # A reset drops the queues the count was claimed against, so it must drop + # the count too or the leak shrinks the budget process-wide. + reset_llama_admission_queues() + revived = get_llama_admission_queue("http://llama.test:1") + fresh = revived.reserve(capacity = 1, config = config).lease_nowait() + assert fresh.park(), "reset leaked the park count" + fresh.release() + + asyncio.run(scenario()) + + +def test_the_park_budget_leaves_the_executor_room_to_work(monkeypatch): + # The pool already permits `capacity` pending prompts and every park admits + # one more, so the budget must account for both. Swept across executor sizes + # rather than read off this host, since a container gets a small one. + for cpus in (1, 2, 4, 8, 16, 28, 64): + workers = min(32, cpus + 4) + monkeypatch.setattr(llama_admission, "_executor_workers", lambda w = workers: w) + reserve = llama_admission._executor_reserve(workers) + assert reserve >= 2, f"{workers} workers left no reserve" + + # Even the smallest executor fits the two simultaneous prompts #7455 needs. + assert llama_admission._max_parked(1) >= 2, f"no room for two on {workers} workers" + assert llama_admission._max_parked(1) <= workers // 2 + # A backend whose --parallel alone fills the executor gets no parks. + assert llama_admission._max_parked(workers) == 0 + for capacity in range(0, workers + 8): + budget = llama_admission._max_parked(capacity) + assert budget >= 0, f"negative budget at capacity {capacity}" + assert ( + budget == 0 or capacity + budget <= workers - reserve + ), f"{workers} workers: capacity {capacity} plus {budget} parks leaves no room" + + +def test_the_park_budget_follows_the_executors_own_cpu_count(monkeypatch): + # 3.13 sizes ThreadPoolExecutor from process_cpu_count(), which honours CPU + # affinity and cgroup quotas; cpu_count() would budget from the whole host + # inside a one-core container. Pulled apart here, since they usually match. + import concurrent.futures + + monkeypatch.setattr(os, "cpu_count", lambda: 64) + if hasattr(os, "process_cpu_count"): + monkeypatch.setattr(os, "process_cpu_count", lambda: 1) + # Against the real thing rather than the formula: the default executor is a + # plain ThreadPoolExecutor(), so its own sizing is the answer on any version. + with concurrent.futures.ThreadPoolExecutor() as pool: + assert llama_admission._executor_workers() == pool._max_workers + + +def test_the_stream_retries_a_park_that_was_refused(): + # _park_admission short-circuits on `on == _parked`, so recording a refused + # park as parked would skip every later approval in the run even once the + # budget frees up. Structural because that only shows on a second approval. + import ast + + # Read rather than import: routes.inference pulls in the whole app. + route = os.path.join(_backend, "routes", "inference.py") + with open(route, encoding = "utf-8") as handle: + tree = ast.parse(handle.read()) + helpers = [ + node + for node in ast.walk(tree) + if isinstance(node, ast.AsyncFunctionDef) and node.name == "_park_admission" + ] + assert len(helpers) == 1, f"expected one _park_admission, found {len(helpers)}" + + guards = [ + node + for node in ast.walk(helpers[0]) + if isinstance(node, ast.If) + and isinstance(node.test, ast.UnaryOp) + and isinstance(node.test.op, ast.Not) + and isinstance(node.test.operand, ast.Call) + and getattr(node.test.operand.func, "attr", None) == "park" + and getattr(node.test.operand.func.value, "id", None) == "lease" + ] + assert len(guards) == 1, "lease.park()'s answer is ignored" + assert all( + isinstance(stmt, ast.Return) for stmt in guards[0].body + ), "a refused park must leave _parked alone, so a later approval retries it" + + +def test_the_park_budget_counts_every_live_backend(monkeypatch): + # base_url takes a fresh port on every load, so a reload mints a queue while + # the old one drains. Prompts on both park threads of the one executor, so a + # budget sized from either backend alone lets them add up past the reserve. + monkeypatch.setattr(llama_admission, "_executor_workers", lambda: 32) + + async def scenario(): + config = LlamaAdmissionConfig() + old = get_llama_admission_queue("http://llama.test:1") + draining = old.reserve(capacity = 16, config = config).lease_nowait() + assert draining is not None # in flight, so the registry keeps this queue + + new = get_llama_admission_queue("http://llama.test:2") + lease = new.reserve(capacity = 16, config = config).lease_nowait() + assert lease is not None + + # 16 slots each against 32 workers: their prompts alone can fill it. + assert llama_admission._max_parked(16) > 0, "this test needs a budget to remove" + assert not lease.park(), "budget sized from one backend of two" + + draining.release() # the old backend drains and is up for eviction + assert lease.park(), "an idle backend still counted against the budget" + lease.release() + + asyncio.run(scenario()) + + +def test_the_park_budget_is_freed_when_the_prompt_is_answered(monkeypatch): + # The executor thread comes back the moment the answer arrives, before the + # resume queues for a slot. Holding the budget until the slot lands refuses + # someone else's park, and that someone holds the slot the resumer wants. + monkeypatch.setattr(llama_admission, "_executor_workers", lambda: 32) + + async def scenario(): + config = LlamaAdmissionConfig() + queue = get_llama_admission_queue("http://llama.test") + + parked = [] + for _ in range(llama_admission._max_parked(1)): + lease = queue.reserve(capacity = 1, config = config).lease_nowait() + assert lease is not None and lease.park() + parked.append(lease) + + blocked = queue.reserve(capacity = 1, config = config).lease_nowait() + assert blocked is not None + assert not blocked.park(), "the budget was not full to begin with" + + # One prompt is answered. Its slot is taken, so the resume queues for one. + resumed = asyncio.ensure_future(parked[0].unpark_async(poll_s = 0.01)) + await asyncio.sleep(0.05) + assert not resumed.done(), "the resume needs to still be waiting for its slot" + + assert blocked.park(), "budget held for a prompt wait that is over" + # Which is what frees the slot the resumer was waiting for. + await asyncio.wait_for(resumed, timeout = 2) + for lease in parked[1:] + [blocked]: + lease.release() + parked[0].release() + + asyncio.run(scenario()) + + +def test_releasing_a_parked_holder_returns_its_budget(monkeypatch): + # A client that disconnects on the prompt releases straight out of parked, + # never unparking. Its executor thread went with it, so keeping the budget + # would lose one for the life of the process. + monkeypatch.setattr(llama_admission, "_executor_workers", lambda: 32) + + async def scenario(): + config = LlamaAdmissionConfig() + queue = get_llama_admission_queue("http://llama.test") + + parked = [] + for _ in range(llama_admission._max_parked(1)): + lease = queue.reserve(capacity = 1, config = config).lease_nowait() + assert lease is not None and lease.park() + parked.append(lease) + + blocked = queue.reserve(capacity = 1, config = config).lease_nowait() + assert blocked is not None + assert not blocked.park(), "the budget was not full to begin with" + + parked[0].release() + assert blocked.park(), "a released park never gave its budget back" + for lease in parked[1:] + [blocked]: + lease.release() + + asyncio.run(scenario()) From 767f2f36fbbab3ff29bcb4ca74347f973c93afa0 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 14:54:01 -0700 Subject: [PATCH 02/16] Windows setup: route the stale-manifest failure through Exit-SetupFailure (#7569) The manifest-removal guard added in #7492 exits with a bare 'exit 1', so in Tauri mode the installer never emits the [TAURI:ERROR] line and the desktop UI falls back to a generic failure instead of naming the cause. Every other failure path in studio/setup.ps1 goes through Exit-SetupFailure, and tests/sh/test_tauri_retry_failure_context.sh asserts that invariant, so 'Repo tests (CPU)' has been red on main since that merge. Co-authored-by: danielhanchen --- studio/setup.ps1 | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/studio/setup.ps1 b/studio/setup.ps1 index 0734b9c2fa..a4eb54a9ef 100644 --- a/studio/setup.ps1 +++ b/studio/setup.ps1 @@ -3122,7 +3122,7 @@ sys.exit(0 if install_manifest.remove_manifest() else 1) if (-not $_ManifestDropped) { Write-Host "[ERROR] Could not remove the stale unsloth_install_manifest.json." -ForegroundColor Red Write-Host " Refusing to install behind a marker that still reports this venv as complete." -ForegroundColor Red - exit 1 + Exit-SetupFailure "Could not remove the stale unsloth_install_manifest.json" } if ($script:UnslothVerbose) { From e662af769bfacd5755449e87fd62855ec86f3680 Mon Sep 17 00:00:00 2001 From: JoshuaL3000 Date: Wed, 29 Jul 2026 06:38:57 +0800 Subject: [PATCH 03/16] fix: enable XPU support and update hardcoded CUDA selections for tests (#7401) * fix: add XPU device support and update hardcoded CUDA selections * fix: add XPU device support for pytest CUDA skipped tests * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Fix device handling for PR #7401 - perplexity_eval.py: use DEVICE_TYPE_TORCH, not DEVICE_TYPE. The latter can be "hip" or "mlx", which .to() rejects, so this regressed ROCm. - test_batched_leftpad_generation_gpu.py: XPU diverges here today, so mark it non-strict xfail on XPU instead of reverting to a CUDA-only guard. Keeps the real XPU gap visible and turns green once it is fixed. - Guard torch.xpu.is_available() with hasattr, matching device_type.py. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Re-enable the flash varlen attention test in CI for PR #7401 attention_dispatch.py now predefines flash_attn_func / flash_attn_varlen_func as None, so test_run_attention_flash_varlen_receives_window_and_softcap no longer needs flash_attn importable to be monkeypatched. Verified on a runner shaped like the CPU-only one: the test fails against main's attention_dispatch and passes at this head, so the deselect is now dead weight. * Tighten comments for PR #7401 Drop the hasattr rationale: torch.xpu has existed since torch 2.3 and the dependency floor is 2.4, so no supported build predates the namespace. The guard stays as cheap defence, but the comment claimed something untrue. --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: danielhanchen --- .github/workflows/consolidated-tests-ci.yml | 10 +++--- .../test_merge_model_perplexity_llama-3.2.py | 11 +++---- .../test_merge_model_perplexity_mistral.py | 11 +++---- .../test_merge_model_perplexity_phi_4.py | 11 +++---- ...st_merged_model_perplexity_llama-3.1-8b.py | 11 +++---- .../test_merged_model_perplexity_qwen_2.5.py | 13 +++----- tests/test_fp8_tiny_e8m0.py | 10 +++--- tests/utils/perplexity_eval.py | 5 ++- .../test_batched_leftpad_generation_gpu.py | 17 ++++++++-- tests/utils/test_packing.py | 20 +++++++++--- tests/utils/test_qat.py | 8 ++++- tests/utils/test_rope_scaling_drift.py | 32 ++++++++++--------- unsloth/utils/attention_dispatch.py | 2 ++ 13 files changed, 94 insertions(+), 67 deletions(-) diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml index 489ee4ca08..c75880fa72 100644 --- a/.github/workflows/consolidated-tests-ci.yml +++ b/.github/workflows/consolidated-tests-ci.yml @@ -372,12 +372,10 @@ jobs: tests/python/test_fast_language_model_text_only.py \ tests/test_bad_mappings_redirect.py \ tests/test_prefetch_snapshot_scope.py \ - tests/test_gemma_2b_mapper_key.py \ - --deselect 'tests/utils/test_attention_masks.py::test_run_attention_flash_varlen_receives_window_and_softcap' - # The deselected test monkeypatches flash_attn_varlen_func, which is - # only bound on the module when `flash_attn` is importable. flash_attn - # requires CUDA + dev toolchain, which the CPU-only ubuntu-latest - # runner does not have. The other Bucket-A tests pass cleanly. + tests/test_gemma_2b_mapper_key.py + # test_run_attention_flash_varlen_receives_window_and_softcap was deselected + # until attention_dispatch.py predefined flash_attn_varlen_func as None; it + # monkeypatches that name, so it no longer needs flash_attn on this runner. - name: unsloth_zoo @ ${{ env.UNSLOTH_ZOO_REF }} — full pytest (CPU) # 106 of 111 test_* in unsloth_zoo are CPU-only. The two CUDA-skip diff --git a/tests/saving/language_models/test_merge_model_perplexity_llama-3.2.py b/tests/saving/language_models/test_merge_model_perplexity_llama-3.2.py index 3b75a13756..a549e58562 100644 --- a/tests/saving/language_models/test_merge_model_perplexity_llama-3.2.py +++ b/tests/saving/language_models/test_merge_model_perplexity_llama-3.2.py @@ -96,12 +96,11 @@ def load_and_compute_8bit_ppl( if __name__ == "__main__": mp.set_start_method("spawn", force = True) - if torch.cuda.is_bf16_supported(): - compute_dtype = torch.bfloat16 - attn_implementation = "flash_attention_2" - else: - compute_dtype = torch.float16 - attn_implementation = "sdpa" + from unsloth import is_bfloat16_supported + from unsloth.models._utils import HAS_FLASH_ATTENTION + + compute_dtype = torch.bfloat16 if is_bfloat16_supported() else torch.float16 + attn_implementation = "flash_attention_2" if HAS_FLASH_ATTENTION else "sdpa" model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/Llama-3.2-3B-Instruct", diff --git a/tests/saving/language_models/test_merge_model_perplexity_mistral.py b/tests/saving/language_models/test_merge_model_perplexity_mistral.py index 8cc833c2b1..50b0d3caf4 100644 --- a/tests/saving/language_models/test_merge_model_perplexity_mistral.py +++ b/tests/saving/language_models/test_merge_model_perplexity_mistral.py @@ -121,12 +121,11 @@ def load_and_compute_8bit_ppl( if __name__ == "__main__": mp.set_start_method("spawn", force = True) - if torch.cuda.is_bf16_supported(): - compute_dtype = torch.bfloat16 - attn_implementation = "flash_attention_2" - else: - compute_dtype = torch.float16 - attn_implementation = "sdpa" + from unsloth import is_bfloat16_supported + from unsloth.models._utils import HAS_FLASH_ATTENTION + + compute_dtype = torch.bfloat16 if is_bfloat16_supported() else torch.float16 + attn_implementation = "flash_attention_2" if HAS_FLASH_ATTENTION else "sdpa" model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/mistral-7b-v0.3", diff --git a/tests/saving/language_models/test_merge_model_perplexity_phi_4.py b/tests/saving/language_models/test_merge_model_perplexity_phi_4.py index 6f79bfdb71..9c7f6c77af 100644 --- a/tests/saving/language_models/test_merge_model_perplexity_phi_4.py +++ b/tests/saving/language_models/test_merge_model_perplexity_phi_4.py @@ -98,12 +98,11 @@ def load_and_compute_8bit_ppl( if __name__ == "__main__": mp.set_start_method("spawn", force = True) - if torch.cuda.is_bf16_supported(): - compute_dtype = torch.bfloat16 - attn_implementation = "flash_attention_2" - else: - compute_dtype = torch.float16 - attn_implementation = "sdpa" + from unsloth import is_bfloat16_supported + from unsloth.models._utils import HAS_FLASH_ATTENTION + + compute_dtype = torch.bfloat16 if is_bfloat16_supported() else torch.float16 + attn_implementation = "flash_attention_2" if HAS_FLASH_ATTENTION else "sdpa" model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/Phi-4", diff --git a/tests/saving/language_models/test_merged_model_perplexity_llama-3.1-8b.py b/tests/saving/language_models/test_merged_model_perplexity_llama-3.1-8b.py index c07b37024f..dcbaad13e1 100644 --- a/tests/saving/language_models/test_merged_model_perplexity_llama-3.1-8b.py +++ b/tests/saving/language_models/test_merged_model_perplexity_llama-3.1-8b.py @@ -95,12 +95,11 @@ def load_and_compute_8bit_ppl( if __name__ == "__main__": mp.set_start_method("spawn", force = True) - if torch.cuda.is_bf16_supported(): - compute_dtype = torch.bfloat16 - attn_implementation = "flash_attention_2" - else: - compute_dtype = torch.float16 - attn_implementation = "sdpa" + from unsloth import is_bfloat16_supported + from unsloth.models._utils import HAS_FLASH_ATTENTION + + compute_dtype = torch.bfloat16 if is_bfloat16_supported() else torch.float16 + attn_implementation = "flash_attention_2" if HAS_FLASH_ATTENTION else "sdpa" model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/Llama-3.1-8B-Instruct", diff --git a/tests/saving/language_models/test_merged_model_perplexity_qwen_2.5.py b/tests/saving/language_models/test_merged_model_perplexity_qwen_2.5.py index cb444d1591..cfa364c697 100644 --- a/tests/saving/language_models/test_merged_model_perplexity_qwen_2.5.py +++ b/tests/saving/language_models/test_merged_model_perplexity_qwen_2.5.py @@ -164,12 +164,11 @@ def load_and_compute_8bit_ppl( if __name__ == "__main__": mp.set_start_method("spawn", force = True) - if torch.cuda.is_bf16_supported(): - compute_dtype = torch.bfloat16 - attn_implementation = "flash_attention_2" - else: - compute_dtype = torch.float16 - attn_implementation = "sdpa" + from unsloth import is_bfloat16_supported + from unsloth.models._utils import HAS_FLASH_ATTENTION + + compute_dtype = torch.bfloat16 if is_bfloat16_supported() else torch.float16 + attn_implementation = "flash_attention_2" if HAS_FLASH_ATTENTION else "sdpa" model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/Qwen2.5-7B-Instruct", @@ -210,8 +209,6 @@ if __name__ == "__main__": loftq_config = None, ) - from unsloth import is_bfloat16_supported - trainer = SFTTrainer( model = model, tokenizer = tokenizer, diff --git a/tests/test_fp8_tiny_e8m0.py b/tests/test_fp8_tiny_e8m0.py index cf49c8c92f..df40879d5a 100644 --- a/tests/test_fp8_tiny_e8m0.py +++ b/tests/test_fp8_tiny_e8m0.py @@ -11,7 +11,11 @@ dequant reference. import pytest import torch -pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason = "needs CUDA") +cuda_available = torch.cuda.is_available() +xpu_available = hasattr(torch, "xpu") and torch.xpu.is_available() +dev = "cuda" if cuda_available else "xpu" if xpu_available else "cpu" + +pytestmark = pytest.mark.skipif(not (cuda_available or xpu_available), reason = "needs CUDA or XPU") def _reference(X, weight, scale, block): @@ -27,7 +31,6 @@ def test_tiny_non_tileable_forward_backward_matches_reference(): from unsloth.kernels.fp8 import FP8BlockQuantLinear torch.manual_seed(0) - dev = "cuda" block = [128, 128] m, n = 8, 8 # non-tileable, in-dim % 128 != 0 weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16) # (out=m, in=n) @@ -50,7 +53,6 @@ def test_e8m0_scale_is_upcast_and_runs(): if not hasattr(torch, "float8_e8m0fnu"): pytest.skip("torch build lacks float8_e8m0fnu") - dev = "cuda" m, n = 8, 8 weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16) scale = (torch.rand(1, 1, device = dev) + 1.0).to(torch.float8_e8m0fnu) @@ -70,7 +72,6 @@ def test_rectangular_block_dequant_matches_reference(): from unsloth.kernels.fp8 import _blockwise_weight_dequant_any_shape torch.manual_seed(0) - dev = "cuda" block = [64, 128] m, n = 64, 256 # evenly tiled: 64 % 64 == 0, 256 % 128 == 0 weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16) @@ -94,7 +95,6 @@ def test_e8m0_scale_preserves_non_default_block_size_attr(): pytest.skip("torch build lacks float8_e8m0fnu") torch.manual_seed(0) - dev = "cuda" block = [64, 64] # in-dim 96 is not divisible by block[1]=64 -> forward takes the torch dequant # fallback (no fp8 matmul kernel). Scale shape (2, 2) validates for [64, 64] but diff --git a/tests/utils/perplexity_eval.py b/tests/utils/perplexity_eval.py index 5f33a24d53..cdd30e5511 100644 --- a/tests/utils/perplexity_eval.py +++ b/tests/utils/perplexity_eval.py @@ -2,6 +2,9 @@ from tqdm import tqdm import torch import pandas as pd +# DEVICE_TYPE_TORCH, not DEVICE_TYPE: the latter can be "hip"/"mlx", which .to() rejects. +from unsloth.device_type import DEVICE_TYPE_TORCH + model_comparison_results = {} @@ -17,7 +20,7 @@ def ppl_model(model, tokenizer, dataset): for begin_loc in range(0, seq_len, stride): end_loc = min(begin_loc + max_length, seq_len) trg_len = end_loc - prev_end_loc - input_ids = encodings.input_ids[:, begin_loc:end_loc].to("cuda") + input_ids = encodings.input_ids[:, begin_loc:end_loc].to(DEVICE_TYPE_TORCH) target_ids = input_ids.clone() target_ids[:, :-trg_len] = -100 pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0 diff --git a/tests/utils/test_batched_leftpad_generation_gpu.py b/tests/utils/test_batched_leftpad_generation_gpu.py index df03125bc2..13db22461e 100644 --- a/tests/utils/test_batched_leftpad_generation_gpu.py +++ b/tests/utils/test_batched_leftpad_generation_gpu.py @@ -4,7 +4,7 @@ Greedy generation in a left-padded batch must match solo batch-size-1 generation for the first PREFIX_TOKENS tokens (the bug makes padded rows diverge into garbage immediately; a full-length match would be flaky due to benign batch-numerics tie-flips deep in the sequence) and must not be -gibberish. Skipped without CUDA. Run: `python -m pytest +gibberish. Skipped without a GPU. Run: `python -m pytest tests/utils/test_batched_leftpad_generation_gpu.py -v`. """ @@ -12,8 +12,19 @@ import pytest import torch cuda_available = torch.cuda.is_available() +xpu_available = hasattr(torch, "xpu") and torch.xpu.is_available() +device = "cuda" if cuda_available else "xpu" if xpu_available else "cpu" -pytestmark = pytest.mark.skipif(not cuda_available, reason = "requires a CUDA GPU") +# Non-strict rather than CUDA-only: keeps the XPU divergence visible, and goes +# green by itself once XPU generation is fixed. +pytestmark = [ + pytest.mark.skipif(not (cuda_available or xpu_available), reason = "requires a CUDA or XPU GPU"), + pytest.mark.xfail( + xpu_available and not cuda_available, + reason = "batched left-padded generation diverges on XPU", + strict = False, + ), +] MODEL_NAME = "unsloth/Qwen2.5-0.5B-Instruct" MAX_NEW_TOKENS = 32 @@ -53,7 +64,7 @@ def _chat(tokenizer, prompt): def _generate(model, tokenizer, texts): inputs = tokenizer(texts, return_tensors = "pt", padding = True, add_special_tokens = False).to( - "cuda" + device ) with torch.inference_mode(): out = model.generate( diff --git a/tests/utils/test_packing.py b/tests/utils/test_packing.py index 1b8bb65058..0be3018cde 100644 --- a/tests/utils/test_packing.py +++ b/tests/utils/test_packing.py @@ -44,6 +44,8 @@ def _build_packed_training_setup(tmp_path, device): dtype = torch.bfloat16 else: dtype = torch.float16 + elif device.type == "xpu": + dtype = torch.bfloat16 try: model, tokenizer = FastLanguageModel.from_pretrained( @@ -76,8 +78,8 @@ def _build_packed_training_setup(tmp_path, device): max_length = 64, logging_steps = 1, max_steps = 1, - fp16 = device.type == "cuda" and not torch.cuda.is_bf16_supported(), - bf16 = device.type == "cuda" and torch.cuda.is_bf16_supported(), + fp16 = dtype == torch.float16, + bf16 = dtype == torch.bfloat16, dataset_num_proc = 1, output_dir = str(tmp_path), packing = True, @@ -974,7 +976,12 @@ def test_enable_sample_packing(): def test_enable_sample_packing_trl_collator(tmp_path): - device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + elif torch.xpu.is_available(): + device = torch.device("xpu") + else: + device = torch.device("cpu") model, _, trainer, _ = _build_packed_training_setup(tmp_path, device) enable_sample_packing(model, trainer) @@ -1030,7 +1037,12 @@ def test_enable_padding_free_metadata(): def test_packing_sdpa(tmp_path): - device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + elif torch.xpu.is_available(): + device = torch.device("xpu") + else: + device = torch.device("cpu") model, batch, trainer, llama_mod = _build_packed_training_setup(tmp_path, device) assert "packed_seq_lengths" in batch diff --git a/tests/utils/test_qat.py b/tests/utils/test_qat.py index 79d955164f..0b942d5c32 100644 --- a/tests/utils/test_qat.py +++ b/tests/utils/test_qat.py @@ -130,8 +130,14 @@ def _test_fake_quantizers_are_called( # Weight fake quantizers must always be called. assert child.weight_fake_quantizer.count == 1 + if torch.cuda.is_available(): + device = torch.device("cuda") + elif torch.xpu.is_available(): + device = torch.device("xpu") + else: + pytest.skip("No GPU available") for k, v in example_inputs.items(): - example_inputs[k] = v.cuda() + example_inputs[k] = v.to(device) model.apply(_swap_fake_quantizers) model(**example_inputs) model.apply(_assert_fake_quantizers_are_called) diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py index eba89734f7..7fe4e74d5c 100644 --- a/tests/utils/test_rope_scaling_drift.py +++ b/tests/utils/test_rope_scaling_drift.py @@ -15,18 +15,20 @@ import pytest import torch -def _has_real_cuda(): - try: - torch.zeros(1).to("cuda") - return True - except Exception: - return False +def _has_real_gpu(): + for backend in ("cuda", "xpu"): + try: + torch.zeros(1).to(backend) + return True + except Exception: + pass + return False -HAS_REAL_CUDA = _has_real_cuda() -requires_cuda = pytest.mark.skipif( - not HAS_REAL_CUDA, - reason = "LlamaRotaryEmbedding builds per-device CUDA caches in __init__", +HAS_REAL_GPU = _has_real_gpu() +requires_gpu = pytest.mark.skipif( + not HAS_REAL_GPU, + reason = "LlamaRotaryEmbedding builds per-device caches in __init__ (needs CUDA or XPU)", ) REPO_ROOT = Path(__file__).resolve().parents[2] @@ -360,7 +362,7 @@ def _cos_at_position(rot, position): # --- Layer 3: CUDA behavioral guard (real instantiation needs a device) --- -@requires_cuda +@requires_gpu def test_constructor_applies_llama3_scaling(): config = _make_config(LLAMA3_ROPE_SCALING) rot = _unsloth_rotary(config) @@ -371,7 +373,7 @@ def test_constructor_applies_llama3_scaling(): ), "LlamaRotaryEmbedding built from a llama3 config produced unscaled inv_freq (issue #2405)." -@requires_cuda +@requires_gpu def test_constructor_unscaled_config_uses_vanilla_inv_freq(): rot = _unsloth_rotary(_make_config(None)) got = rot.inv_freq.float().cpu() @@ -381,7 +383,7 @@ def test_constructor_unscaled_config_uses_vanilla_inv_freq(): ), "LlamaRotaryEmbedding with no rope_scaling must use the vanilla inv_freq" -@requires_cuda +@requires_gpu def test_cos_cache_differs_between_scaled_and_unscaled_at_long_position(): scaled = _unsloth_rotary(_make_config(LLAMA3_ROPE_SCALING)) unscaled = _unsloth_rotary(_make_config(None)) @@ -397,7 +399,7 @@ def test_cos_cache_differs_between_scaled_and_unscaled_at_long_position(): ) -@requires_cuda +@requires_gpu def test_extended_cache_keeps_scaling_after_growth(): scaled = _unsloth_rotary(_make_config(LLAMA3_ROPE_SCALING)) # Grow past the initial cache size (mirrors long-context decode). @@ -456,7 +458,7 @@ def _build_longrope_rotary(): return rot, config -@requires_cuda +@requires_gpu @pytest.mark.parametrize( "build", [_build_llama3_rotary, _build_longrope_rotary], ids = ["llama3", "longrope"] ) diff --git a/unsloth/utils/attention_dispatch.py b/unsloth/utils/attention_dispatch.py index eda6103d5b..54f8100ca1 100644 --- a/unsloth/utils/attention_dispatch.py +++ b/unsloth/utils/attention_dispatch.py @@ -31,6 +31,8 @@ from ..utils.packing import ( build_xformers_block_causal_mask, ) +flash_attn_func = None +flash_attn_varlen_func = None if HAS_FLASH_ATTENTION: from flash_attn import flash_attn_func, flash_attn_varlen_func HAS_XFORMERS = xformers is not None From 036fa6009538548ce70426d3ad42e6092ac6f067 Mon Sep 17 00:00:00 2001 From: Nilay <118994073+NilayYadav@users.noreply.github.com> Date: Wed, 29 Jul 2026 04:11:47 +0530 Subject: [PATCH 04/16] Studio: pass raise_on_error=False on the stdio MCP call path (#7517) --- studio/backend/core/inference/mcp_client.py | 7 ++- .../backend/tests/test_mcp_flatten_result.py | 43 +++++++++++++++++++ .../backend/tests/test_mcp_stdio_sessions.py | 40 +++++++++++++---- 3 files changed, 81 insertions(+), 9 deletions(-) diff --git a/studio/backend/core/inference/mcp_client.py b/studio/backend/core/inference/mcp_client.py index 0256df944e..98112c6d5b 100644 --- a/studio/backend/core/inference/mcp_client.py +++ b/studio/backend/core/inference/mcp_client.py @@ -971,7 +971,12 @@ def _call_stdio_tool( raise RuntimeError("MCP server connection is not available") else: rem = _remaining() - coro = _race_tool_call(session.client.call_tool(name, args), rem, cancel_event) + # raise_on_error=False for the same reason as the one-shot path. + coro = _race_tool_call( + session.client.call_tool(name, args, raise_on_error = False), + rem, + cancel_event, + ) return session.run(coro, rem) except (_MCPCancelled, asyncio.TimeoutError): # _race_tool_call cancels the pending call but cancellation is diff --git a/studio/backend/tests/test_mcp_flatten_result.py b/studio/backend/tests/test_mcp_flatten_result.py index 7daee799f9..618c5ccfe6 100644 --- a/studio/backend/tests/test_mcp_flatten_result.py +++ b/studio/backend/tests/test_mcp_flatten_result.py @@ -175,3 +175,46 @@ def test_call_tool_sync_passes_raise_on_error_false_and_keeps_error_images(monke assert out.startswith("Error: boom") assert MCP_IMAGES_SENTINEL in out assert is_tool_error(out) + + +def test_stdio_session_call_also_passes_raise_on_error_false(monkeypatch): + seen = {} + + class _FakeStdioClient: + def __init__(self): + self.connected = False + self.transport = SimpleNamespace(_is_session_dead = lambda: False) + + async def __aenter__(self): + self.connected = True + return self + + async def __aexit__(self, *exc): + self.connected = False + + def is_connected(self): + return self.connected + + async def call_tool( + self, + name, + args, + raise_on_error = True, + ): + seen["raise_on_error"] = raise_on_error + return _result(_text("boom"), _image(), is_error = True) + + monkeypatch.setattr( + mcp_client, "_client", lambda url, headers, use_oauth = False: _FakeStdioClient() + ) + try: + out = call_tool_sync( + "npx fake-stdio-server", None, "take_screenshot", {}, scope = "s=p:t=thread1" + ) + finally: + mcp_client.close_stdio_sessions() + + assert seen["raise_on_error"] is False + assert out.startswith("Error: boom") + assert MCP_IMAGES_SENTINEL in out + assert is_tool_error(out) diff --git a/studio/backend/tests/test_mcp_stdio_sessions.py b/studio/backend/tests/test_mcp_stdio_sessions.py index d714d9d640..37c812677a 100644 --- a/studio/backend/tests/test_mcp_stdio_sessions.py +++ b/studio/backend/tests/test_mcp_stdio_sessions.py @@ -60,7 +60,12 @@ class FakeClient: def is_connected(self) -> bool: return self.connected - async def call_tool(self, name: str, args: dict): + async def call_tool( + self, + name: str, + args: dict, + raise_on_error: bool = True, + ): if self.call_delay: await asyncio.sleep(self.call_delay) if self.fail_next: @@ -120,10 +125,15 @@ def test_tool_error_does_not_recycle_session(fake_clients, monkeypatch): from fastmcp.exceptions import ToolError class ToolFailure(FakeClient): - async def call_tool(self, name, args): + async def call_tool( + self, + name, + args, + raise_on_error = True, + ): if name == "boom": raise ToolError("tool exploded") # tool-level: session stays connected - return await super().call_tool(name, args) + return await super().call_tool(name, args, raise_on_error) monkeypatch.setattr( mcp_client, "_client", lambda url, headers, use_oauth = False: ToolFailure(url) @@ -441,12 +451,17 @@ def test_overlapping_calls_serialize_on_shared_session(fake_clients, monkeypatch active = 0 max_active = 0 - async def call_tool(self, name, args): + async def call_tool( + self, + name, + args, + raise_on_error = True, + ): OverlapDetect.active += 1 OverlapDetect.max_active = max(OverlapDetect.max_active, OverlapDetect.active) try: await asyncio.sleep(0.2) - return await super().call_tool(name, args) + return await super().call_tool(name, args, raise_on_error) finally: OverlapDetect.active -= 1 @@ -473,9 +488,14 @@ def test_timeout_budget_spans_connect_and_call(fake_clients, monkeypatch): await asyncio.sleep(0.4) return await super().__aenter__() - async def call_tool(self, name, args): + async def call_tool( + self, + name, + args, + raise_on_error = True, + ): await asyncio.sleep(0.5) - return await super().call_tool(name, args) + return await super().call_tool(name, args, raise_on_error) monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: SlowBoth(url)) start = time.monotonic() @@ -565,7 +585,11 @@ def test_execute_tool_config_check_tracks_row(tmp_path, monkeypatch): def test_multi_block_result_flattens_through_session(fake_clients): - async def _rich_call(name, args): + async def _rich_call( + name, + args, + raise_on_error = True, + ): return SimpleNamespace( content = [ SimpleNamespace(type = "text", text = "### Page"), From 31969053d8caf3baae51dcc515acfa76d096afae Mon Sep 17 00:00:00 2001 From: Vineeth Sai Varikuntla Date: Tue, 28 Jul 2026 16:13:00 -0700 Subject: [PATCH 05/16] Cover the FP8 row-scaling path in the newer-mapper probe (#7516) * Pin the newer-mapper FP8 probe with tests that can fail The two identity assertions added in #7478 compare the returned FP8 tables against the installed ones, but the fixture serves the same mapper.py as both the installed and the fetched source and exec always allocates fresh dicts, so they pin allocation rather than provenance and hold for any new dict. Replace them with two tests that drive get_model_name end to end: one splices an FP8 entry into the fetched source only and asserts the upgrade error still fires, the other serves a mapper.py with no FP8 tables and asserts the 4bit half of the probe survives, which is the regression #7497 fixed. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Tighten the resolver stub for PR #7516 - Restore the fp8_block/fp8_row identity assert alongside the new provenance test. It is weak, not vacuous: it still catches a probe that hands back the installed table objects, and it costs nothing to keep. - Bind Version and transformers_version in the stub namespace. Both are unreached under the current gates, so a change to either would fail with a bare NameError instead of the assertion. Merged main, which clears the unrelated test_runtime_text_encoding failure the branch inherited from its base. * Cover the FP8 row-scaling path instead of duplicating the block one The two tests this PR originally added were already covered by tests/test_new_mapper_fetched_fp8.py from #7497. An 8-mutant matrix over loader_utils.py found nothing they caught that the existing file did not, so they are dropped and test_new_mapper_no_global_leak.py goes back to main. Two real gaps were open, both on the row branch that load_in_fp8 = True plus UNSLOTH_HAS_FBGEMM selects ahead of block: - the FBGEMM row branch in __get_model_name could be deleted outright with every test still green - _resolve_with_mappers could ignore its fp8_row argument and silently fall back to the installed row table Adds two tests to the existing file, reusing its _load_resolver rather than a second harness. The row-only fixture splices into the fetched row table alone, since an entry the block table also knows lets the block branch answer and masks the regression. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: danielhanchen --- tests/test_new_mapper_fetched_fp8.py | 53 ++++++++++++++++++++++++++++ 1 file changed, 53 insertions(+) diff --git a/tests/test_new_mapper_fetched_fp8.py b/tests/test_new_mapper_fetched_fp8.py index 2835aadb59..bdd1b241fa 100644 --- a/tests/test_new_mapper_fetched_fp8.py +++ b/tests/test_new_mapper_fetched_fp8.py @@ -14,6 +14,11 @@ Two gaps it misses: future rename): reading them with ``[]`` raises ``KeyError`` into the bare ``except``, taking the 4bit half, the probe's whole purpose, down with it. +Both of the above only reach the block table. The last two tests take the row branch, which +``load_in_fp8 = True`` plus ``UNSLOTH_HAS_FBGEMM`` selects ahead of block: deleting that branch, +or dropping ``_resolve_with_mappers``' ``fp8_row`` argument so it falls back to the installed +table, both leave every other test here green. + ``loader_utils`` imports torch, so ast-extract the resolvers and run them against a stubbed ``requests``, as in ``tests/test_bad_mappings_redirect.py``. """ @@ -33,6 +38,8 @@ _NEW_OFFICIAL = "zeta-org/Zeta-9B-Only-On-Main-FP8" _NEW_BLOCK = "unsloth/Zeta-9B-Only-On-Main-FP8-Block" _NEW_ROW = "unsloth/Zeta-9B-Only-On-Main-FP8-Row" _ANCHOR = ' "unsloth/Kimi-K2-Instruct-BF16" : (' +# Row table only, so the block branch cannot answer for it and mask a row-path regression. +_ROW_ONLY = "zeta-org/Zeta-9B-Row-Only-FP8" def _mapper_source(): @@ -51,6 +58,11 @@ def _with_extra_fp8_model(source): return source.replace(_ANCHOR, entry + _ANCHOR, 1) +def _with_row_only_fp8_model(source): + """Fetched row table only. Block must not know it, or the block branch answers instead.""" + return source + f'\nFLOAT_TO_FP8_ROW_MAPPER["{_ROW_ONLY.lower()}"] = "{_NEW_ROW}"\n' + + def _without_fp8_tables(source): """A mapper.py from before the fp8 tables existed.""" return source.replace("FLOAT_TO_FP8_BLOCK_MAPPER", "SOME_OTHER_BLOCK_TABLE").replace( @@ -153,3 +165,44 @@ def test_probe_survives_a_fetched_mapper_without_the_fp8_tables(monkeypatch): assert ( int_to_float and float_to_int and map_to_16bit ), "a fetched mapper.py without the fp8 tables must not take the 4bit upgrade check down" + + +def test_fbgemm_prefers_the_row_table_over_the_block_one(monkeypatch): + """With FBGEMM, `load_in_fp8 = True` must resolve row-scaled, not blockwise.""" + monkeypatch.setenv("UNSLOTH_HAS_FBGEMM", "1") + namespace = _load_resolver(_mapper_source()) + row = namespace["FLOAT_TO_FP8_ROW_MAPPER"] + block = namespace["FLOAT_TO_FP8_BLOCK_MAPPER"] + + key = next(k for k in row if k in block and row[k] != block[k]) + resolved = namespace["get_model_name"](key, load_in_4bit = False, load_in_fp8 = True) + + assert resolved == row[key], ( + f"FBGEMM must take the row branch for {key!r}, got {resolved!r} " + f"(the blockwise answer is {block[key]!r})" + ) + + +def test_probe_answers_for_a_row_only_repo_the_fetched_mapper_knows(monkeypatch): + """The row half of the probe needs the FETCHED row table, same as the block half.""" + monkeypatch.setenv("UNSLOTH_HAS_FBGEMM", "1") + installed = _mapper_source() + namespace = _load_resolver(installed) + installed_row = namespace["FLOAT_TO_FP8_ROW_MAPPER"] + key = _ROW_ONLY.lower() + assert key not in installed_row, "the installed row table must not know it" + assert key not in namespace["FLOAT_TO_FP8_BLOCK_MAPPER"], "no block entry, or block answers" + + _install_fake_requests(monkeypatch, _with_row_only_fp8_model(installed)) + _install_fake_vllm_absent(monkeypatch, namespace) + + try: + resolved = namespace["get_model_name"](_ROW_ONLY, load_in_4bit = False, load_in_fp8 = True) + except NotImplementedError as error: + assert "not supported in your current Unsloth version" in str(error) + else: + raise AssertionError( + f"a fetched-only row-scaled repo must raise the upgrade error, got {resolved!r}" + ) + + assert namespace["FLOAT_TO_FP8_ROW_MAPPER"] is installed_row From 7ac75c6572421acb86fbb35d38ab686dec61729a Mon Sep 17 00:00:00 2001 From: Vineeth Sai Varikuntla Date: Tue, 28 Jul 2026 17:40:43 -0700 Subject: [PATCH 06/16] Parse a .json dataset file as one JSON document instead of line-by-line (#7422) --------- Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com> --- .github/workflows/consolidated-tests-ci.yml | 3 +- tests/test_raw_text_json_loading.py | 128 ++++++++++++++++++++ unsloth/dataprep/raw_text.py | 40 ++++-- 3 files changed, 162 insertions(+), 9 deletions(-) create mode 100644 tests/test_raw_text_json_loading.py diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml index c75880fa72..afad1b6c46 100644 --- a/.github/workflows/consolidated-tests-ci.yml +++ b/.github/workflows/consolidated-tests-ci.yml @@ -372,7 +372,8 @@ jobs: tests/python/test_fast_language_model_text_only.py \ tests/test_bad_mappings_redirect.py \ tests/test_prefetch_snapshot_scope.py \ - tests/test_gemma_2b_mapper_key.py + tests/test_gemma_2b_mapper_key.py \ + tests/test_raw_text_json_loading.py # test_run_attention_flash_varlen_receives_window_and_softcap was deselected # until attention_dispatch.py predefined flash_attn_varlen_func as None; it # monkeypatches that name, so it no longer needs flash_attn on this runner. diff --git a/tests/test_raw_text_json_loading.py b/tests/test_raw_text_json_loading.py new file mode 100644 index 0000000000..27e636da18 --- /dev/null +++ b/tests/test_raw_text_json_loading.py @@ -0,0 +1,128 @@ +"""Regression test for .json parsing in unsloth/dataprep/raw_text.py. + +Both .json and .jsonl map to the "json_lines" handler, which used to parse the +file one line at a time. A real .json file is a single JSON document (commonly +a top-level list of records), so every line failed json.loads, the whole +document was dropped, and the handler returned "" (load_from_file then rejected +the valid file as "empty"). The handler now parses the file as one JSON value +first and falls back to line-by-line for true .jsonl. + +raw_text.py's only third-party import is `datasets`, so we stub it and exec the +module directly, with no `import unsloth` (which needs a GPU / unsloth_zoo). +""" + +import json +import sys +import types +from pathlib import Path + +RAW_TEXT_PATH = Path(__file__).parents[1] / "unsloth" / "dataprep" / "raw_text.py" + + +def _load_raw_text(): + sys.modules.setdefault("datasets", types.SimpleNamespace(Dataset = object)) + module = types.ModuleType("unsloth_raw_text_under_test") + exec( + compile(RAW_TEXT_PATH.read_text(encoding = "utf-8"), str(RAW_TEXT_PATH), "exec"), + module.__dict__, + ) + return module + + +def test_json_document_is_parsed_whole(tmp_path): + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + path = tmp_path / "data.json" + path.write_text( + json.dumps([{"text": "hello world"}, {"text": "second sample"}], indent = 2), encoding = "utf-8" + ) + assert loader._read_file_by_format(str(path), "json_lines") == "hello world\n\nsecond sample" + + +def test_jsonl_is_still_parsed_line_by_line(tmp_path): + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + path = tmp_path / "data.jsonl" + path.write_text('{"text": "a"}\n{"text": "b"}\n', encoding = "utf-8") + assert loader._read_file_by_format(str(path), "json_lines") == "a\n\nb" + + +def test_jsonl_is_never_materialized(tmp_path): + """A .jsonl file must keep streaming, whole-document parsing is only for .json.""" + real_open = open + + class _StreamOnlyFile: + """File wrapper that fails the test if the whole file is pulled into memory.""" + + def __init__(self, handle): + self.handle = handle + + def __enter__(self): + return self + + def __exit__(self, *exc_info): + self.handle.close() + return False + + def __iter__(self): + return iter(self.handle) + + def read(self, *args, **kwargs): + raise AssertionError(".jsonl was read whole instead of streamed line by line") + + def seek(self, *args, **kwargs): + raise AssertionError(".jsonl was re-read instead of streamed line by line") + + module = _load_raw_text() + module.open = lambda *args, **kwargs: _StreamOnlyFile(real_open(*args, **kwargs)) + + path = tmp_path / "big.jsonl" + path.write_text('{"text": "a"}\n\n{"text": "b"}\nnot json at all\n', encoding = "utf-8") + loader = module.RawTextDataLoader(tokenizer = object()) + assert loader._read_file_by_format(str(path), "json_lines") == "a\n\nb" + + +def test_json_holding_json_lines_still_falls_back(tmp_path): + """A .json file that actually holds JSON Lines still parses, via the per-line fallback.""" + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + path = tmp_path / "mislabelled.json" + path.write_text('{"text": "a"}\n{"text": "b"}\n', encoding = "utf-8") + assert loader._read_file_by_format(str(path), "json_lines") == "a\n\nb" + + +def test_utf8_bom_json_document_is_parsed(tmp_path): + """Windows tooling prefixes a UTF-8 BOM; it must not sink the whole document.""" + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + path = tmp_path / "bom.json" + path.write_text( + json.dumps([{"text": "hello world"}, {"text": "second sample"}], indent = 2), + encoding = "utf-8-sig", + ) + assert path.read_bytes().startswith(b"\xef\xbb\xbf") + assert loader._read_file_by_format(str(path), "json_lines") == "hello world\n\nsecond sample" + + +def test_utf8_bom_jsonl_keeps_first_record(tmp_path): + """A BOM must not silently drop the first .jsonl record.""" + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + path = tmp_path / "bom.jsonl" + path.write_text('{"text": "a"}\n{"text": "b"}\n', encoding = "utf-8-sig") + assert loader._read_file_by_format(str(path), "json_lines") == "a\n\nb" + + +def test_utf8_bom_json_holding_json_lines_falls_back(tmp_path): + """The per-line fallback re-reads from byte 0, so the BOM must be stripped again.""" + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + path = tmp_path / "bom_mislabelled.json" + path.write_text('{"text": "a"}\n{"text": "b"}\n', encoding = "utf-8-sig") + assert loader._read_file_by_format(str(path), "json_lines") == "a\n\nb" + + +def test_utf8_bom_plain_text_and_csv(tmp_path): + """The BOM also leaks into .txt training text and the first .csv column name.""" + loader = _load_raw_text().RawTextDataLoader(tokenizer = object()) + txt = tmp_path / "bom.txt" + txt.write_text("hello", encoding = "utf-8-sig") + assert loader._read_file_by_format(str(txt), "plain_text") == "hello" + + csv_path = tmp_path / "bom.csv" + csv_path.write_text("text,other\nhello,x\n", encoding = "utf-8-sig") + assert loader._read_file_by_format(str(csv_path), "csv_text_column") == "hello" diff --git a/unsloth/dataprep/raw_text.py b/unsloth/dataprep/raw_text.py index fdaba181f1..0920e2d7f4 100644 --- a/unsloth/dataprep/raw_text.py +++ b/unsloth/dataprep/raw_text.py @@ -216,19 +216,32 @@ class RawTextDataLoader: def _read_file_by_format(self, file_path, file_format): """Read file content based on detected format.""" - with open(file_path, "r", encoding = "utf-8") as f: + # utf-8-sig: Windows tooling (PowerShell's Out-File, Excel's "CSV UTF-8") prepends + # a BOM that plain utf-8 keeps as a leading character. Without a BOM it decodes + # exactly like utf-8. + with open(file_path, "r", encoding = "utf-8-sig") as f: if file_format == "plain_text" or file_format == "markdown": return f.read() elif file_format == "json_lines": - lines = [] - for line in f: + if Path(file_path).suffix.lower() == ".json": + # A .json file is a single JSON document (commonly a list + # of records), so parsing it per line drops the whole file. try: - data = json.loads(line.strip()) - text = self._extract_text_from_json(data) - if text: - lines.append(text) + parsed = json.load(f) + records = parsed if isinstance(parsed, list) else [parsed] except json.JSONDecodeError: - continue + # Some files carry JSON Lines under a .json name. + f.seek(0) + records = self._iter_json_lines(f) + else: + # A .jsonl file is one JSON value per line: stay streaming so + # a large file is never held in memory all at once. + records = self._iter_json_lines(f) + lines = [] + for data in records: + text = self._extract_text_from_json(data) + if text: + lines.append(text) return "\n\n".join(lines) elif file_format == "csv_text_column": reader = csv.DictReader(f) @@ -244,6 +257,17 @@ class RawTextDataLoader: _TEXT_FIELDS = ("text", "content", "message", "body", "description", "prompt") _TEXT_COLUMNS = _TEXT_FIELDS + def _iter_json_lines(self, handle): + """Yield one parsed JSON value per line, skipping blank and malformed lines.""" + for line in handle: + line = line.strip() + if not line: + continue + try: + yield json.loads(line) + except json.JSONDecodeError: + continue + def _extract_text_from_json(self, data): """Extract text from JSON object using common field names.""" # Skip non-object lines (str/list/number): `field in data` would be a From 150b5ba25ad22c194da9fa21158b543f0dda3fda Mon Sep 17 00:00:00 2001 From: Kirelos Namroud <87078943+knamroud@users.noreply.github.com> Date: Wed, 29 Jul 2026 03:03:28 +0200 Subject: [PATCH 07/16] feat(studio): adjustable llama-server parallel slots from the web UI (#7447) * feat(studio): share the llama-server --parallel bounds as PARALLEL_MIN/MAX The per-load parallel-slots field needs the same 1..64 range the CLI flag validates, but models/inference.py cannot import run.py (run.py builds the app that imports routes that import models). Promote the bounds into this dependency-free module, which already owns the -np/--parallel semantics, and record the deliberate mirrors that cannot import it (run.py, the unsloth CLI, the web UI). The denylist entry stays: the first-class field is now the single write path for the slot count, so a pass-through would still desync the committed bookkeeping from llama-server. * feat(studio): note the per-load override in the --parallel help text --parallel is now the server-wide default that a per-load n_parallel (the Studio Parallel Slots run setting) can override, not the definitive slot count. Point at the new control so a user does not conclude a restart is the only way to change slots, and record the shared PARALLEL_MIN/MAX mirror alongside the existing CLI one. * feat(studio): add n_parallel to LoadRequest and echo the slot counts LoadRequest.n_parallel (optional, PARALLEL_MIN..PARALLEL_MAX) lets a load pick its own llama-server --parallel count; omitted, the server-wide launch default applies. ValidateModelRequest carries it too so the training-coexistence estimate sizes the KV cache like the follow-up load rather than passing on a smaller footprint. LoadResponse and InferenceStatusResponse gain both requested_parallel_slots (what the load was invoked with) and parallel_slots (what llama-server actually runs after the fitter's slot reduction), so a client can tell an honored request from a reduced one. Both are None where --parallel has no meaning: non-GGUF loads and the diffusion runner. * feat(studio): record the requested parallel-slot count on the backend The auto GPU-memory fit may launch fewer slots than requested to keep the model fully on GPU, so the committed effective count cannot answer "is the live server what this request asked for?". Store the invoked count separately (mirroring the _requested_n_ctx pattern) from the pre-reduction pending kwargs, expose it as requested_parallel_slots, and have _already_in_target_state compare requested-vs-requested: comparing against the effective count would reload -- and re-reduce -- forever on an identical Apply. The comparison sits in the non-diffusion branch, since the diffusion runner ignores --parallel entirely. The requested value shares the effective count's lifecycle, so every unload/kill path clears it and a stale count cannot poison the next load's dedupe. * feat(studio): honor a per-load parallel-slot count in /load and /validate Resolve the slot count once per load -- the request field if set, else the server-wide launch default -- and feed it to every consumer that must agree: the training-coexistence guard, the llama-server load kwargs, and the reload dedupe. Without the dedupe comparison a changed slot count would be swallowed as already_loaded; it compares requested-vs-requested and skips the diffusion runner, which ignores --parallel. app.state.llama_parallel_slots is deliberately never written: it stays the launch intent and the admission-queue fallback, so one load's override cannot leak into later loads. /validate resolves the same way so its estimate cannot undercount what the load then allocates. Both /load returns and /status echo the counts through one helper, which reports None for diffusion -- its load never commits a count, so echoing the reset placeholder would fabricate an "invoked with 1 slot". * feat(studio): accept nParallel in the chat-preset load config ChatPresetLoadConfig is extra="forbid", so a preset carrying the new parallel slots knob would 422 the whole settings sync without this field. Bounds come from the shared PARALLEL_MIN/MAX rather than literals, so a future range change cannot start rejecting presets the UI still allows. * test(studio): cover the per-load parallel-slots knob Pins the behaviors a regression would silently break: the requested-vs-effective dedupe (comparing against the reduced count would reload forever), the diffusion skip and its None echo, the requested count's reset lifecycle, and its commit from the pre-reduction pending kwargs. Also pins the three bounds mirrors that cannot import PARALLEL_MIN/MAX (run.py, the unsloth CLI, the web UI) plus the preset model that can, so a range change cannot leave one of them clamping or rejecting at the old limit. * test(studio): refresh the --parallel denylist comments for the UI knob The pinned rationale said the typer flag owns the slot count and pointed users at a Studio restart. Parallel Slots / LoadRequest.n_parallel is now the other managed writer, and the 1..64 guard is the shared PARALLEL_MIN/MAX -- a reader following the old comments would conclude the UI control does not exist. * feat(studio): note the per-load override in the CLI --parallel help Both the plain-serve and `unsloth studio run` flags now describe a server-wide default the Studio Parallel Slots run setting can override per load, matching the backend help text. * feat(studio): remember a per-model Parallel Slots override nParallel joins the per-model config with the same null-means-follow-the-default convention as the other knobs: null keeps the server-wide --parallel count, so a blank control never pins a number and isDefaultConfig still deletes an otherwise-untouched config instead of storing it. The value is re-clamped to N_PARALLEL_MIN/MAX on every localStorage read and write (the store is user-editable), and listing it in STORED_CONFIG_FIELDS keeps it from being dropped as an unknown key. Legacy blobs predate the knob, so their migration carries null. No schema-version bump: an additive optional field, like the GPU fields before it. * feat(studio): bridge nParallel between the per-model config and the store The config->store, store->config and equality helpers all need the new field: without the equality arm a slots-only edit reads as unchanged, so Apply is dropped and the dirty state never lights up. * feat(studio): track the parallel-slot override in the chat runtime store nParallel holds the editable override and loadedNParallel the value the last successful load sent, which the failed-switch rollback re-sends. Both are per-model: they clear on unload and on a model switch, unlike the standing preferences (GPU memory mode, speculative type) that survive one. There is deliberately no backend-echo field for the control: the echo is the resolved count, so adopting it would pin a blank "follow the server default" input to an explicit number. * feat(studio): type n_parallel and the slot-count echoes The load request gains the optional per-load slot count, and both the load response and the status payload gain requested_parallel_slots (invoked) and parallel_slots (actually running after the fitter's reduction). Keys stay snake_case: the payload is serialized as-is, with no case conversion. * feat(studio): forward n_parallel to the validate preflight validateModel builds its own body rather than forwarding the load payload, so the slot count has to be listed explicitly. Slots scale the KV estimate, and the preflight exists to refuse a load the training guard would then 409 -- an unforwarded count would validate a smaller footprint than the load allocates. * feat(studio): include nParallel in the active model's config The sidebar assembles the active model's config from individually subscribed store fields; an unsubscribed field would leave the form showing a stale value after any external change. * feat(studio): add the Parallel Slots control to the run settings A numeric input in the GGUF advanced section, blank meaning "follow the server default". It clamps on change like the Draft Tokens field rather than using NumericValueInput, so there is no blur-draft to lose when the user types a value and immediately clicks Load. hasNonDefaultAdvanced counts it too, so a remembered override reopens the advanced section instead of hiding the setting that is actually in effect. * feat(studio): key the sidebar config form on nParallel too The signature drives the remount that re-seeds the form; without the new field an externally changed slot count would leave the sidebar showing the old one. * feat(studio): send the Parallel Slots override on load performLoad snapshots the slot count at click time (staged run-settings config first, else the store) and sends it on both the validate preflight and the load, so the two size the same footprint. A cross-model switch re-baselines it like the other per-model knobs -- the previous model's count must not follow onto the next one -- and the failed-switch rollback re-sends the previous model's value so a rescue reload cannot silently drop to the server default. The success path keeps the click-time value rather than the response echo: the echo is the count the fitter resolved, so adopting it would turn a blank "follow the server default" control into an explicit pin. Slots are GGUF-only, so a transformers load sends and records null instead of a phantom override. * feat(studio): carry the slot override through the compare-pane load The compare pane builds its own load request, so it needs the field explicitly or a pane with a remembered override would load at the server default. Its validate preflight sends the same count, matching the comment above it that promises validation is sized exactly as the load below. GGUF-gated on both calls, and the store adopts the pane's own click-time value rather than the resolved echo, mirroring the single-model path. * feat(studio): honor the remembered slot override on startup auto-load The auto-load path reads the per-model config and forwards every other remembered knob, so a remembered Parallel Slots value was the one setting lost on the "load last used model" path: llama-server came back at the server-wide default with the control showing blank, and the first manual Apply afterwards then forced a needless reload because the counts disagreed. * feat(studio): seed the slot baseline from the status echo Only the rollback baseline is seeded, never the editable control: the echo is the resolved count, so adopting it would pin a blank "follow the server default" input to a number. Without the seed, loadedNParallel stayed null after a tab reload or a second tab adopting the running model, and a failed switch then rolled the previous model back at the server default while every other knob was restored. * feat(studio): capture Parallel Slots in chat presets The knob joins the preset load config end to end: captured from the store, re-clamped when read back (persisted presets are untrusted input), applied on switch, and summarized in the preset chip. Its default is null, so coalesceDefaultLoadKnobs keeps a default-only preset empty rather than persisting a no-op override. * feat(studio): re-derive the preset state when Parallel Slots changes Both preset memos snapshot the store through capturePresetLoadConfig, so without the new dependency a slots-only edit left the unsaved-changes flag and the load summary showing the previous value. * test(studio): pin the Parallel Slots wiring end to end Source-contract coverage for the hops a refactor can silently drop: the three /load builders (interactive, compare pane, startup auto-load) and their validate preflights, per-model persistence and clamping, the UI row, and the status seed -- including the negative assertion that hydration seeds only the rollback baseline, never the control, so the resolved echo cannot pin a blank "server default" input. * test(studio): pin nParallel in the preset load config Covers capture, clamped read-back and apply on the frontend, plus the backend field itself: ChatPresetLoadConfig is extra="forbid", so a missing or drifted field 422s every settings sync that carries a preset. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Fall back to one slot when llama-server lacks --kv-unified for PR #7447 Without --kv-unified an explicit --parallel N makes llama-server give each slot -c/N, so on a build without the flag choosing N slots silently shrinks every context window for a feature that build cannot serve. Clamp to one slot and log why, placed after the requested count is captured so the echo still reports it and before the KV estimates so the fit matches what actually launches. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Clear the slot control on load paths that never send it, and size the training guard for diffusion Four review findings on the per-load Parallel Slots knob. The editable nParallel control means "follow the server default" when null, so any success path that does not send a slot count has to clear it. Three paths kept a value staged for a different model: - chat-adapter.ts, cached non-GGUF auto-load: the interactive and compare builders already clear both fields for a non-GGUF response, this third one did not. The field never renders for a non-GGUF target, so the stale count was invisible and unclearable from the UI yet still persisted, and it flips isDefaultConfig so a user with no overrides silently gets a stored entry. - chat-adapter.ts, fresh-model fallback: its request omits n_parallel but its success state resynced every other knob and left the slots alone, so a staged edit survived against a server running the default and the next Apply reloaded at a count that load never sent. - apply-inference-status-to-store.ts: on a model change underneath the tab every sibling knob adopts the new model's status, but nParallel updated only its baseline, so the previous model's explicit count followed onto the new model and saving or reloading there pinned it. Clear the control and keep seeding the baseline for the rollback. The training-coexistence guard sized a diffusion GGUF with the requested slot count. _estimate_kv_cache_bytes scales the SWA cache with slots (swa_limit = swa * slots + ubatch), but load_model hands a diffusion target to _start_diffusion_server before the slot plumbing, so that runner is always single-slot. At the new default of 4 this inflated the estimate and could 409 a load that fits. An unclassified GGUF keeps the requested count. Backend base KV depends on -c alone, not on --parallel, which is why only the SWA term is affected: llama.cpp PR 14363 and discussion 4130. Tests: three training-guard cases in test_parallel_slots_per_load.py and one source contract in test_model_picker_contracts.py, each mutation-checked. 174 passed across the backend slot/admission/training suites, 56 across the frontend contract suites. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Keep the slot control when re-adopting the running model, and never record slots for a diffusion load Two follow-ups from the latest review round. The first is a regression from c796393. That commit cleared the slot control whenever hydratingExistingModel was set, to stop model A's count following onto model B. But that flag is also set on the resident-model adopt path: when the store checkpoint is an external provider id and the user re-picks the still loaded local model, applyActiveModelStatusToStore is called with the external id as previousCheckpoint, so the flag is unconditionally true. The clear then wiped the config applyPerModelConfigToRuntime had restored two lines earlier, and it was the only knob that did, because the siblings re-adopt the status echo while this one cleared. Gate the clear on the tab's own baseline no longer matching the running count: a genuine A to B swap still clears, re-adopting the same model keeps its value. The second revises an earlier call of mine. I rejected the diffusion phantom as cosmetic because the backend ignores the value on every send. The sharpened report is right and my rejection was wrong: capturePresetLoadConfig records nParallel with no model gate, a Preset carries no model id, and applying one writes nParallel for whatever model is current. So a count recorded against a diffusion model, which the backend never applied, rides a saved preset onto a text GGUF and becomes a real override the user never chose. Record slots only when the load actually committed them, on all three load builders. Tests: two source contracts in test_model_picker_contracts.py, both mutation checked. Frontend typecheck clean, 58 passed across the contract and preset suites. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Clear the slot baseline when status reports a model without slots Hydrating from a GGUF to a slotless model left loadedNParallel at the previous model's count: the seed only runs when the echo is non-null, and the control clear added earlier touches nParallel alone. The stale baseline is what a failed-switch rollback re-sends, and preset capture reads it, so it could claim slots for a model that never used them. Clear it when status describes a model that cannot have slots. /status omits the echo entirely for non-GGUF and sends an explicit null for the diffusion runner, so keying on is_gguf === false or an explicit null covers both while an absent field on a GGUF, which is how an older backend reports one, still leaves the baseline alone. Test mutation checked; frontend typecheck clean against a fresh npm ci. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Distinguish a same-model re-adopt from a model swap, and size the training guard at the slots that launch for PR #7447 * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Keep the blank slot control across a failed-switch rollback for PR #7447 * Restore a remembered slot override when hydrating a fresh store for PR #7447 * Tighten comments for PR #7447 * Restore a remembered slot override on a model switch too for PR #7447 * Tighten comments and docstrings for PR #7447 * Take the rollback slot intent from the picker's pre-switch snapshot for PR #7447 * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: danielhanchen Co-authored-by: danielhanchen --- studio/backend/core/inference/llama_cpp.py | 23 + .../core/inference/llama_server_args.py | 11 +- studio/backend/models/inference.py | 57 ++ studio/backend/routes/chat_history.py | 2 + studio/backend/routes/inference.py | 71 ++- studio/backend/run.py | 3 +- .../backend/tests/test_llama_server_args.py | 14 +- .../tests/test_parallel_slots_per_load.py | 517 ++++++++++++++++++ .../src/features/chat/api/chat-adapter.ts | 19 + .../src/features/chat/api/chat-api.ts | 2 + .../src/features/chat/chat-settings-sheet.tsx | 3 + .../chat/hooks/use-chat-model-runtime.ts | 46 +- .../lib/apply-inference-status-to-store.ts | 51 ++ .../chat/presets/preset-load-config.ts | 17 + .../src/features/chat/shared-composer.tsx | 12 + .../chat/stores/chat-runtime-store.ts | 10 + .../frontend/src/features/chat/types/api.ts | 17 + .../components/model-config-page.tsx | 41 ++ .../components/sidebar-model-config.tsx | 1 + .../hooks/use-active-model-config.ts | 3 + .../model-config/apply-per-model-config.ts | 3 + .../model-config/per-model-config.ts | 15 + tests/studio/test_chat_preset_load_config.py | 15 + tests/studio/test_model_picker_contracts.py | 247 +++++++++ unsloth_cli/commands/studio.py | 6 +- 25 files changed, 1186 insertions(+), 20 deletions(-) create mode 100644 studio/backend/tests/test_parallel_slots_per_load.py diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py index f76afce9f4..4f2d8cd54a 100644 --- a/studio/backend/core/inference/llama_cpp.py +++ b/studio/backend/core/inference/llama_cpp.py @@ -2181,6 +2181,8 @@ class LlamaCppBackend: self._effective_context_length: Optional[int] = None self._max_context_length: Optional[int] = None self._effective_parallel_slots: int = 1 + # --parallel the last load asked for, before any fit-time reduction. + self._requested_n_parallel: int = 1 self._chat_template: Optional[str] = None self._chat_template_override: Optional[str] = None self._supports_reasoning: bool = False @@ -2417,6 +2419,17 @@ class LlamaCppBackend: slots = 1 return max(1, slots) + @property + def requested_parallel_slots(self) -> int: + """--parallel the last load asked for, before any fit-time reduction. + The reload dedupe compares requested-vs-requested (like requested_n_ctx); + the effective count would reload forever after a fitter reduction.""" + try: + slots = int(getattr(self, "_requested_n_parallel", 1)) + except (TypeError, ValueError): + slots = 1 + return max(1, slots) + @property def max_context_length(self) -> Optional[int]: """Return the largest context that fits on this hardware at load time. @@ -2442,6 +2455,8 @@ class LlamaCppBackend: def _reset_effective_parallel_slots(self) -> None: self._effective_parallel_slots = 1 + # Cleared with the effective count so a stale value can't skew the dedupe. + self._requested_n_parallel = 1 @staticmethod def _read_rss_bytes(pid: int) -> Optional[int]: @@ -6787,6 +6802,7 @@ class LlamaCppBackend: chat_template_override = chat_template_override, extra_args = extra_args, is_vision = is_vision, + n_parallel = n_parallel, preserve_multi_gpu_on_layer = preserve_multi_gpu_on_layer, ): logger.info( @@ -9066,6 +9082,8 @@ class LlamaCppBackend: self._extra_args = list(extra_args) self._extra_args_source = (model_identifier, hf_variant) self._requested_n_ctx = int(n_ctx) + # Local n_parallel may have been reduced above; the snapshot has the ask. + self._requested_n_parallel = max(1, int(_pending_load_kwargs["n_parallel"])) # Commit the known-good snapshot + whether MTP+tensor is live, then # watch this load for a mid-generation crash. self._last_load_kwargs = _pending_load_kwargs @@ -9478,6 +9496,7 @@ class LlamaCppBackend: tensor_split: Optional[List[float]] = None, gpu_ids: Optional[List[int]] = None, mtp_draft_path: Optional[str] = None, + n_parallel: int = 1, preserve_multi_gpu_on_layer: bool = False, ) -> bool: """True iff the live server already satisfies these load kwargs. @@ -9542,6 +9561,10 @@ class LlamaCppBackend: # A GPU-memory-mode flip (Unsloth / manual) must always reload. if self._gpu_memory_mode != gpu_memory_mode: return False + # Requested-vs-requested (like n_ctx): comparing the effective count + # would reload forever whenever the fitter launched fewer slots. + if self._requested_n_parallel != max(1, int(n_parallel)): + return False # Manual: a layer-count change always reloads (covers Auto(-1) <-> a # pinned count); MoE/split only matter with an explicit offload. if gpu_memory_mode == "manual" and ( diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py index 2ecd7e3e2e..7391e62516 100644 --- a/studio/backend/core/inference/llama_server_args.py +++ b/studio/backend/core/inference/llama_server_args.py @@ -16,11 +16,18 @@ from __future__ import annotations import os from typing import Iterable, Mapping, Optional +# Valid llama-server --parallel range, shared with LoadRequest.n_parallel. +# Mirrored by callers that cannot import this: run.py and unsloth_cli/commands/ +# studio.py (_PARALLEL_MIN/MAX), per-model-config.ts (N_PARALLEL_MIN/MAX); +# test_parallel_slots_per_load.py pins them together. +PARALLEL_MIN = 1 +PARALLEL_MAX = 64 + # Each group = every alias (short + long) of one hard-denied flag. # Extend the matching group when llama.cpp adds a new alias. _DENYLIST_GROUPS: tuple[frozenset[str], ...] = ( - # Parallel slots: owned by typer --parallel; a pass-through would desync - # app.state.llama_parallel_slots from llama-server. + # Parallel slots: owned by typer --parallel and LoadRequest.n_parallel; a + # pass-through would desync the slot bookkeeping from llama-server. frozenset({"-np", "--parallel", "--n-parallel"}), # Model identity: Unsloth resolves it from LoadRequest; a second -m would # load a different model than Unsloth thinks it loaded. diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py index acd60dd0b9..0edd1aa37f 100644 --- a/studio/backend/models/inference.py +++ b/studio/backend/models/inference.py @@ -18,6 +18,7 @@ from pydantic import ( model_validator, ) +from core.inference.llama_server_args import PARALLEL_MAX, PARALLEL_MIN from picker.schemas import MAX_CHAT_TEMPLATE_BYTES @@ -113,6 +114,18 @@ class LoadRequest(BaseModel): "'mtp' or 'mtp+ngram'." ), ) + n_parallel: Optional[int] = Field( + None, + ge = PARALLEL_MIN, + le = PARALLEL_MAX, + description = ( + "Parallel decode slots for llama-server (--parallel) for this " + f"load ({PARALLEL_MIN}..{PARALLEL_MAX}). Omit for the server-wide " + "default set at launch (the --parallel CLI flag). The VRAM fitter " + "may launch fewer slots to keep the model fully on GPU. Ignored " + "for non-GGUF models." + ), + ) tensor_parallel: bool = Field( False, description = ( @@ -265,6 +278,16 @@ class ValidateModelRequest(BaseModel): "delegate fitting to llama.cpp, while explicit layers are user-owned." ), ) + n_parallel: Optional[int] = Field( + None, + ge = PARALLEL_MIN, + le = PARALLEL_MAX, + description = ( + "Parallel decode slots intended for the follow-up load, so the " + "coexistence estimate sizes the KV cache like /load. Omit for the " + "server-wide --parallel default." + ), + ) include_context_length: bool = Field( False, description = "Also read the native context length from the local GGUF header. " @@ -533,6 +556,23 @@ class LoadResponse(BaseModel): "or None for automatic selection." ), ) + requested_parallel_slots: Optional[int] = Field( + None, + description = ( + "Parallel decode slots the load was invoked with (per-load " + "n_parallel, else the server-wide --parallel default). None for " + "non-GGUF loads and for the diffusion runner, which ignores " + "--parallel." + ), + ) + parallel_slots: Optional[int] = Field( + None, + description = ( + "Serving slots the active llama-server actually runs (--parallel " + "after any fit-time slot reduction). None for non-GGUF loads and " + "for the diffusion runner, which ignores --parallel." + ), + ) class UnloadResponse(BaseModel): @@ -708,6 +748,23 @@ class InferenceStatusResponse(BaseModel): "or None for automatic selection." ), ) + requested_parallel_slots: Optional[int] = Field( + None, + description = ( + "Parallel decode slots the active load was invoked with (per-load " + "n_parallel, else the server-wide --parallel default). None when " + "no GGUF model is loaded and for the diffusion runner, which " + "ignores --parallel." + ), + ) + parallel_slots: Optional[int] = Field( + None, + description = ( + "Serving slots the active llama-server actually runs (--parallel " + "after any fit-time slot reduction). None when no GGUF model is " + "loaded and for the diffusion runner, which ignores --parallel." + ), + ) llama_cpp_supports_mtp: bool = Field( True, description = ( diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py index aa59716315..4180518837 100644 --- a/studio/backend/routes/chat_history.py +++ b/studio/backend/routes/chat_history.py @@ -11,6 +11,7 @@ from fastapi import APIRouter, Depends, HTTPException, Query, Request from pydantic import BaseModel, ConfigDict, Field, ValidationError from auth.authentication import get_current_subject +from core.inference.llama_server_args import PARALLEL_MAX, PARALLEL_MIN from loggers import get_logger from utils.utils import safe_curated_detail, log_and_http_error from storage.studio_db import ( @@ -169,6 +170,7 @@ class ChatPresetLoadConfig(BaseModel): kvCacheDtype: Optional[str] = None speculativeType: Optional[str] = None specDraftNMax: Optional[int] = Field(default = None, ge = 1, le = 16) + nParallel: Optional[int] = Field(default = None, ge = PARALLEL_MIN, le = PARALLEL_MAX) tensorParallel: Optional[bool] = None gpuMemoryMode: Optional[Literal["manual"]] = None gpuLayers: Optional[int] = None diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py index 53b4136e32..12547277f5 100644 --- a/studio/backend/routes/inference.py +++ b/studio/backend/routes/inference.py @@ -3294,10 +3294,25 @@ def _is_explicit_tensor_drop(request: LoadRequest) -> bool: return override is not None and override.strip().lower() != "tensor" +def _parallel_slot_echo(llama_backend: LlamaCppBackend) -> dict: + """requested/effective parallel-slot fields for /load and /status echoes. + + The diffusion runner ignores ``--parallel`` and never commits a count, so it + reports None like the non-GGUF paths; echoing the reset placeholder 1 would + fabricate an "invoked with 1 slot".""" + if llama_backend.is_diffusion: + return {"requested_parallel_slots": None, "parallel_slots": None} + return { + "requested_parallel_slots": llama_backend.requested_parallel_slots, + "parallel_slots": llama_backend.effective_parallel_slots, + } + + def _request_matches_loaded_settings( request: LoadRequest, llama_backend: LlamaCppBackend, effective_chat_template_override: Optional[str] = None, + requested_parallel_slots: Optional[int] = None, ) -> bool: """True iff every runtime setting on the request matches the loaded server. Caller has already checked model+variant+is_loaded. See #5401. @@ -3306,11 +3321,22 @@ def _request_matches_loaded_settings( launched (user override, else a bundled family template such as the gemma-4 override), so the dedup compares against what the backend actually holds rather than the raw request field. Defaults to the request field for - callers that do not resolve a bundled override.""" + callers that do not resolve a bundled override. + + ``requested_parallel_slots`` is the resolved count the load would use + (per-load ``n_parallel``, else the server-wide default); None skips it.""" # Compare requested n_ctx (not effective) so VRAM-cap doesn't mask an # Auto-vs-explicit slider flip. if request.max_seq_length != llama_backend.requested_n_ctx: return False + # Requested-vs-requested for the same reason: the fitter may launch fewer + # slots. Diffusion ignores --parallel, so a change there must not reload. + if ( + requested_parallel_slots is not None + and not llama_backend.is_diffusion + and int(requested_parallel_slots) != llama_backend.requested_parallel_slots + ): + return False if _normalise_settings_str(request.cache_type_kv) != _normalise_settings_str( llama_backend.cache_type_kv ): @@ -4730,6 +4756,20 @@ def _guard_chat_load_against_training( cpu_only = LlamaCppBackend._effective_gpu_count() == 0, ) + # Size with the count that will actually launch, or a load that fits gets a + # 409: diffusion never receives --parallel, and load_model clamps to 1 on an + # llama-server without --kv-unified. An unclassified GGUF keeps the ask. + if is_gguf and n_parallel > 1: + if diffusion_kind is True: + n_parallel = 1 + else: + try: + caps = LlamaCppBackend.probe_server_capabilities() + if caps.get("found") and not caps.get("supports_kv_unified"): + n_parallel = 1 + except Exception as e: + logger.warning("Could not probe llama-server slots for chat-load guard: %s", e) + required_override_gb = ( _estimate_gguf_required_gb( config, @@ -5272,6 +5312,17 @@ async def _load_model_impl( backend = get_inference_backend() llama_backend = get_llama_cpp_backend() + # Resolve the slot count once (per-load field, else the server-wide + # --parallel default) so the dedupe, the training guard and the load + # kwargs all size against what launches. app.state stays the launch + # intent / admission fallback; getattr because direct callers have no app. + _app_state = getattr(getattr(fastapi_request, "app", None), "state", None) + _n_parallel = ( + request.n_parallel + if request.n_parallel is not None + else getattr(_app_state, "llama_parallel_slots", 1) + ) + is_direct_gguf_request = model_identifier.lower().endswith(".gguf") if request.gguf_variant or is_direct_gguf_request: gguf_variant_matches = is_direct_gguf_request or bool( @@ -5289,6 +5340,7 @@ async def _load_model_impl( request, llama_backend, effective_chat_template_override, + requested_parallel_slots = _n_parallel, ) # Skip if a prior audio probe failed -- let load_model retry. and getattr(llama_backend, "_audio_probed", True) @@ -5343,6 +5395,7 @@ async def _load_model_impl( n_moe_layers = llama_backend.n_moe_layers, gpu_ids = llama_backend.gpu_ids, requested_gpu_ids = llama_backend.requested_gpu_ids, + **_parallel_slot_echo(llama_backend), ) else: if ( @@ -5481,7 +5534,7 @@ async def _load_model_impl( max_seq_length = request.max_seq_length, requested_gpu_ids = effective_gpu_ids, llama_extra_args = extra_llama_args, - n_parallel = getattr(fastapi_request.app.state, "llama_parallel_slots", 1), + n_parallel = _n_parallel, cache_type_kv = request.cache_type_kv, tensor_parallel = bool(request.tensor_parallel), gpu_memory_mode = request.gpu_memory_mode, @@ -5558,7 +5611,6 @@ async def _load_model_impl( # Route to HF or local mode based on config. Run in a thread so the # event loop stays free for progress polling and other requests # during the (potentially long) GGUF download + llama-server start. - _n_parallel = getattr(fastapi_request.app.state, "llama_parallel_slots", 1) # Load kwargs common to HF and local modes; the two differ only by # the model-source args (hf_repo/-token vs gguf_path/mmproj). @@ -5756,6 +5808,7 @@ async def _load_model_impl( n_moe_layers = llama_backend.n_moe_layers, gpu_ids = llama_backend.gpu_ids, requested_gpu_ids = llama_backend.requested_gpu_ids, + **_parallel_slot_echo(llama_backend), ) # ── Standard path: load via Unsloth/transformers ────────── @@ -6156,9 +6209,14 @@ async def validate_model( requested_gpu_ids = effective_gpu_ids, llama_extra_args = effective_extra_args, n_parallel = ( - getattr(fastapi_request.app.state, "llama_parallel_slots", 1) - if fastapi_request is not None - else 1 + request.n_parallel + if request.n_parallel is not None + # Same getattr chain as the load path: preflight must size like the load. + else getattr( + getattr(getattr(fastapi_request, "app", None), "state", None), + "llama_parallel_slots", + 1, + ) ), cache_type_kv = request.cache_type_kv, tensor_parallel = request.tensor_parallel, @@ -6987,6 +7045,7 @@ async def get_status(current_subject: str = Depends(get_current_subject)): n_moe_layers = llama_backend.n_moe_layers, gpu_ids = llama_backend.gpu_ids, requested_gpu_ids = llama_backend.requested_gpu_ids, + **_parallel_slot_echo(llama_backend), llama_cpp_supports_mtp = _supports_mtp, spec_fallback_reason = llama_backend.spec_fallback_reason, llama_cpp_prebuilt_stale = _stale, diff --git a/studio/backend/run.py b/studio/backend/run.py index 8ef1ac06b8..08d1c5299e 100644 --- a/studio/backend/run.py +++ b/studio/backend/run.py @@ -1920,7 +1920,8 @@ def _build_arg_parser(): default = _PARALLEL_DEFAULT_PLAIN, help = ( f"llama-server parallel decode slots ({_PARALLEL_MIN}..{_PARALLEL_MAX}). " - f"Default {_PARALLEL_DEFAULT_PLAIN}." + f"Default {_PARALLEL_DEFAULT_PLAIN}. The Studio run settings " + "(Parallel Slots) override it per load." ), ) return parser diff --git a/studio/backend/tests/test_llama_server_args.py b/studio/backend/tests/test_llama_server_args.py index b2ec5034ac..83934e4130 100644 --- a/studio/backend/tests/test_llama_server_args.py +++ b/studio/backend/tests/test_llama_server_args.py @@ -77,8 +77,7 @@ validate_extra_args = _lsa.validate_extra_args ["--reasoning-format", "deepseek"], ["-rea", "auto"], # Soft-managed: user flags last-wins over Unsloth's auto-set version. - # --parallel / -np / --n-parallel are hard-denied (KV-cache + slot - # count would desync); use `unsloth studio run --parallel N` instead. + # --parallel / -np / --n-parallel are hard-denied; use Parallel Slots. ["-c", "131072"], ["--ctx-size", "8192"], ["--flash-attn", "off"], @@ -128,7 +127,7 @@ def test_non_flag_token_passes_through(): @pytest.mark.parametrize( "denied", [ - # Parallel slots -- owned by the typer --parallel flag. + # Parallel slots -- owned by typer --parallel and LoadRequest.n_parallel. "-np", "--parallel", "--n-parallel", @@ -201,9 +200,8 @@ def test_denylist_rejects_all_aliases(denied): @pytest.mark.parametrize( "args,offending", [ - # Pass-through --parallel would last-wins-override the real slot - # count while Unsloth's KV-cache fit + llama_parallel_slots stay at - # the typer value -- plan vs. process disagree. + # Pass-through --parallel would last-wins-override the real slot count + # while the KV-cache fit and slot bookkeeping stay at the resolved value. (["--parallel", "8"], "--parallel"), (["--parallel=8"], "--parallel"), (["--n-parallel", "16"], "--n-parallel"), @@ -213,7 +211,7 @@ def test_denylist_rejects_all_aliases(denied): # `["-np8"]` must still resolve to managed. (["-np8"], "-np"), (["-np64"], "-np"), - # Out-of-range values that would bypass the typer 1..64 guard. + # Out-of-range values that would bypass the PARALLEL_MIN/MAX bounds. (["--parallel", "999"], "--parallel"), (["-np", "0"], "-np"), (["-np999"], "-np"), @@ -300,7 +298,7 @@ def test_is_managed_flag_true_for_denied(): assert is_managed_flag("--api-key") is True assert is_managed_flag("-m") is True assert is_managed_flag("--model") is True - # Parallel slots owned by the typer --parallel flag. + # Parallel slots owned by typer --parallel and LoadRequest.n_parallel. assert is_managed_flag("--parallel") is True assert is_managed_flag("--n-parallel") is True assert is_managed_flag("-np") is True diff --git a/studio/backend/tests/test_parallel_slots_per_load.py b/studio/backend/tests/test_parallel_slots_per_load.py new file mode 100644 index 0000000000..f4f2d31c6f --- /dev/null +++ b/studio/backend/tests/test_parallel_slots_per_load.py @@ -0,0 +1,517 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +"""Backend contract for the per-load parallel-slots knob. + +An optional ``n_parallel`` (llama-server ``--parallel``) rides on LoadRequest; +omitted, the server-wide launch default (``run.py --parallel``) applies. These +tests pin the pydantic contract and the shared PARALLEL_MIN/MAX mirrors, the +``requested_parallel_slots`` lifecycle, the ``_already_in_target_state`` +requested-vs-requested reload branch with its diffusion skip, and the route +wiring behind the /load, /validate and /status echoes. +""" + +from __future__ import annotations + +import inspect +import re +import struct +import sys +import types as _types +from pathlib import Path + +import pytest + +_BACKEND_DIR = str(Path(__file__).resolve().parent.parent) +if _BACKEND_DIR not in sys.path: + sys.path.insert(0, _BACKEND_DIR) + +# Same external-dep stubs as the other llama_cpp unit tests. +_loggers_stub = _types.ModuleType("loggers") +_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name) +sys.modules.setdefault("loggers", _loggers_stub) + +_structlog_stub = _types.ModuleType("structlog") +_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub") +sys.modules.setdefault("structlog", _structlog_stub) + +# Real httpx: a stub would poison a combined run (routes/inference reads its +# attrs at def time). +import httpx # noqa: F401 + +from core.inference import llama_cpp as llama_cpp_module +from core.inference.llama_server_args import PARALLEL_MAX, PARALLEL_MIN +from core.inference.llama_cpp import LlamaCppBackend +from models.inference import ( + InferenceStatusResponse, + LoadRequest, + LoadResponse, + ValidateModelRequest, +) + + +class _FakeProcess: + def terminate(self): + pass + + def wait(self, timeout = None): + return 0 + + def kill(self): + pass + + def poll(self): + return 0 + + +# ── Pydantic contract ──────────────────────────────────────────────── + + +def test_load_request_defaults_n_parallel_none(): + assert LoadRequest(model_path = "owner/repo").n_parallel is None + + +@pytest.mark.parametrize("value", [PARALLEL_MIN, 4, PARALLEL_MAX]) +def test_load_request_accepts_in_range_n_parallel(value): + assert LoadRequest(model_path = "owner/repo", n_parallel = value).n_parallel == value + + +@pytest.mark.parametrize("value", [0, -1, PARALLEL_MAX + 1]) +def test_load_request_rejects_out_of_range_n_parallel(value): + with pytest.raises(ValueError): + LoadRequest(model_path = "owner/repo", n_parallel = value) + + +def test_load_request_round_trips_json_key(): + req = LoadRequest.model_validate({"model_path": "owner/repo", "n_parallel": 8}) + assert req.n_parallel == 8 + assert req.model_dump()["n_parallel"] == 8 + + +def test_validate_request_n_parallel_contract(): + # /validate sizes like /load, so it carries the same field and bounds. + assert ValidateModelRequest(model_path = "owner/repo").n_parallel is None + assert ( + ValidateModelRequest(model_path = "owner/repo", n_parallel = PARALLEL_MAX).n_parallel + == PARALLEL_MAX + ) + with pytest.raises(ValueError): + ValidateModelRequest(model_path = "owner/repo", n_parallel = PARALLEL_MAX + 1) + + +@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse]) +def test_response_models_emit_parallel_slot_fields(model_cls): + kwargs = ( + dict(status = "loaded", model = "owner/repo", display_name = "repo", inference = {}) + if model_cls is LoadResponse + else {} + ) + empty = model_cls(**kwargs).model_dump() + assert empty["requested_parallel_slots"] is None + assert empty["parallel_slots"] is None + dumped = model_cls(**kwargs, requested_parallel_slots = 8, parallel_slots = 4).model_dump() + assert dumped["requested_parallel_slots"] == 8 + assert dumped["parallel_slots"] == 4 + + +# ── Shared bounds and their deliberate mirrors ─────────────────────── + + +def _mirrored_bounds(source_path: Path) -> tuple[int, int]: + src = source_path.read_text(encoding = "utf-8") + low = re.search(r"^_PARALLEL_MIN\s*=\s*(\d+)$", src, re.MULTILINE) + high = re.search(r"^_PARALLEL_MAX\s*=\s*(\d+)$", src, re.MULTILINE) + assert low and high, f"{source_path} must define _PARALLEL_MIN/_PARALLEL_MAX" + return int(low.group(1)), int(high.group(1)) + + +def test_run_py_mirror_matches_shared_bounds(): + assert _mirrored_bounds(Path(_BACKEND_DIR) / "run.py") == (PARALLEL_MIN, PARALLEL_MAX) + + +def test_cli_mirror_matches_shared_bounds(): + cli = Path(_BACKEND_DIR).parent.parent / "unsloth_cli" / "commands" / "studio.py" + assert _mirrored_bounds(cli) == (PARALLEL_MIN, PARALLEL_MAX) + + +def test_frontend_mirror_matches_shared_bounds(): + # The UI clamps with its own copy; a bumped PARALLEL_MAX that skips it would + # leave the UI silently capping lower. + src = ( + Path(_BACKEND_DIR).parent + / "frontend" + / "src" + / "features" + / "model-picker" + / "model-config" + / "per-model-config.ts" + ).read_text(encoding = "utf-8") + low = re.search(r"^export const N_PARALLEL_MIN = (\d+);$", src, re.MULTILINE) + high = re.search(r"^export const N_PARALLEL_MAX = (\d+);$", src, re.MULTILINE) + assert low and high, "per-model-config.ts must export N_PARALLEL_MIN/MAX" + assert (int(low.group(1)), int(high.group(1))) == (PARALLEL_MIN, PARALLEL_MAX) + + +def test_preset_model_reuses_shared_bounds(): + # Bounds drifting from PARALLEL_MIN/MAX would 422 valid presets on every sync. + from routes.chat_history import ChatPresetLoadConfig + + field = ChatPresetLoadConfig.model_fields["nParallel"] + bounds = {type(m).__name__: getattr(m, "ge", getattr(m, "le", None)) for m in field.metadata} + assert bounds.get("Ge") == PARALLEL_MIN + assert bounds.get("Le") == PARALLEL_MAX + + +# ── requested_parallel_slots lifecycle ─────────────────────────────── + + +@pytest.fixture +def backend(monkeypatch): + monkeypatch.setattr(LlamaCppBackend, "_kill_orphaned_servers", lambda self: 0) + monkeypatch.setattr(llama_cpp_module.atexit, "register", lambda *_args, **_kwargs: None) + return LlamaCppBackend() + + +def test_requested_parallel_slots_initial_value_is_one(backend): + assert backend.requested_parallel_slots == 1 + + +def test_requested_parallel_slots_reflects_field(backend): + backend._requested_n_parallel = 8 + assert backend.requested_parallel_slots == 8 + + +@pytest.mark.parametrize("value", [None, 0, -2, "not-an-int"]) +def test_requested_parallel_slots_invalid_value_falls_back_to_one(backend, value): + backend._requested_n_parallel = value + assert backend.requested_parallel_slots == 1 + + +def test_reset_effective_parallel_slots_also_resets_requested(backend): + backend._requested_n_parallel = 8 + backend._commit_effective_parallel_slots(4) + + backend._reset_effective_parallel_slots() + + assert backend.requested_parallel_slots == 1 + assert backend.effective_parallel_slots == 1 + + +def test_unload_resets_requested_parallel_slots(backend): + backend._process = _FakeProcess() + backend._requested_n_parallel = 8 + + backend.unload_model() + + assert backend.requested_parallel_slots == 1 + + +def test_load_model_commits_requested_from_pending_kwargs(): + # n_parallel may be reduced before the commit, so the requested value must + # come from the pre-reduction pending snapshot. + src = inspect.getsource(LlamaCppBackend.load_model) + commit = src.find( + 'self._requested_n_parallel = max(1, int(_pending_load_kwargs["n_parallel"]))' + ) + healthy = src.find("self._healthy = True\n", 0, commit if commit != -1 else None) + snapshot = src.find("self._last_load_kwargs = _pending_load_kwargs") + assert commit != -1, "load_model must commit the requested slot count" + assert healthy != -1 and healthy < commit < snapshot + + +# ── _already_in_target_state requested-vs-requested branch ─────────── + + +def _loaded_backend() -> LlamaCppBackend: + backend = LlamaCppBackend() + backend._process = _FakeProcess() # is_loaded only checks "is not None" + backend._healthy = True + backend._model_identifier = "owner/repo" + backend._hf_variant = "Q4_K_M" + backend._requested_n_ctx = 8192 + backend._cache_type_kv = None + backend._requested_spec_mode = "auto" + backend._chat_template_override = None + backend._is_vision = False + backend._extra_args = None + backend._gguf_path = None + return backend + + +def _target_state(backend: LlamaCppBackend, n_parallel: int) -> bool: + return backend._already_in_target_state( + gguf_path = None, + model_identifier = "owner/repo", + hf_variant = "Q4_K_M", + n_ctx = 8192, + cache_type_kv = None, + speculative_type = "auto", + chat_template_override = None, + extra_args = None, + is_vision = False, + n_parallel = n_parallel, + ) + + +def test_already_in_target_state_matches_same_slots(): + backend = _loaded_backend() + backend._requested_n_parallel = 4 + assert _target_state(backend, 4) is True + + +def test_already_in_target_state_reloads_on_slots_change(): + backend = _loaded_backend() + backend._requested_n_parallel = 4 + assert _target_state(backend, 8) is False + + +def test_already_in_target_state_compares_requested_not_effective(): + # An identical re-Apply must dedupe even after the fitter reduced the slots. + backend = _loaded_backend() + backend._requested_n_parallel = 8 + backend._commit_effective_parallel_slots(4) + assert _target_state(backend, 8) is True + + +def test_already_in_target_state_ignores_slots_for_diffusion(): + # The diffusion runner ignores --parallel, so a slots change must not reload. + backend = _loaded_backend() + backend._is_diffusion = True + backend._requested_n_parallel = 1 + assert _target_state(backend, 8) is True + + +# ── Route wiring (source contract, mirroring test_gpu_memory_mode) ─── + + +def _route_source() -> str: + return (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8") + + +def _load_impl_source() -> str: + """Body of _load_model_impl only, so positional assertions can't be + satisfied by a later function in the module.""" + src = _route_source() + body = src[src.index("async def _load_model_impl") :] + return body[: body.index("\n@router.")] + + +def test_route_resolves_slots_once_before_dedupe_guard_and_load(): + load_impl = _load_impl_source() + resolve = load_impl.index("request.n_parallel") + fallback = load_impl.index('getattr(_app_state, "llama_parallel_slots", 1)') + dedupe = load_impl.index("requested_parallel_slots = _n_parallel") + guard = load_impl.index("_guard_chat_load_against_training") + # The GGUF launch kwargs, not the guard's own kwarg (which shares the spelling). + load_kwargs = load_impl.index("_common_load_kwargs = dict(") + assert resolve < dedupe, "resolution must precede the reload dedupe" + assert fallback < dedupe + assert resolve < guard < load_kwargs + # Guard and load kwargs share the resolved value; app.state is read once. + assert load_impl.count("n_parallel = _n_parallel") == 2 + assert "n_parallel = _n_parallel" in load_impl[load_kwargs : load_kwargs + 800] + assert load_impl.count('getattr(_app_state, "llama_parallel_slots", 1)') == 1 + # getattr, so a direct caller without an app cannot raise, and no re-read. + assert "fastapi_request.app.state" not in load_impl + + +def test_route_dedupe_compares_requested_slots_and_skips_diffusion(): + match_impl = _route_source()[_route_source().index("def _request_matches_loaded_settings") :] + match_impl = match_impl[: match_impl.index("\ndef ")] + assert "requested_parallel_slots is not None" in match_impl + assert "not llama_backend.is_diffusion" in match_impl + assert "llama_backend.requested_parallel_slots" in match_impl + + +def test_route_echoes_requested_and_effective_slots(): + route_src = _route_source() + # Both /load returns plus the /status GGUF branch, via the shared helper. + assert route_src.count("**_parallel_slot_echo(llama_backend)") == 3 + + +def test_parallel_slot_echo_reports_none_for_diffusion(): + # Diffusion never commits a count, so echoing the reset placeholder 1 would lie. + from routes.inference import _parallel_slot_echo + + backend = _loaded_backend() + backend._requested_n_parallel = 8 + backend._commit_effective_parallel_slots(4) + assert _parallel_slot_echo(backend) == {"requested_parallel_slots": 8, "parallel_slots": 4} + backend._is_diffusion = True + assert _parallel_slot_echo(backend) == { + "requested_parallel_slots": None, + "parallel_slots": None, + } + + +def test_validate_route_prefers_request_n_parallel(): + validate_impl = _route_source()[_route_source().index("async def validate_model") :] + resolve = validate_impl.index("request.n_parallel") + fallback = validate_impl.index('"llama_parallel_slots",') + guard = validate_impl.index("_guard_chat_load_against_training") + assert guard < resolve and guard < fallback, "the guard call resolves the slots inline" + + +def _load_model_source() -> str: + return inspect.getsource(LlamaCppBackend.load_model) + + +def test_slots_fall_back_to_one_without_kv_unified(): + # Without --kv-unified llama-server gives each slot -c/N, so an explicit + # --parallel N shrinks every context window. + src = _load_model_source() + clamp = src.find("supports_kv_unified") + assert clamp != -1, "load_model must check for --kv-unified before honouring the slots" + block = src[clamp : clamp + 700] + assert ( + "n_parallel > 1" in src[clamp - 300 : clamp] + ), "only an explicit multi-slot load is clamped" + assert "n_parallel = 1" in block + + +def test_clamp_sits_between_the_echo_and_the_fit(): + # The echo reports the ask and the fit uses what launches, so the clamp + # belongs between the two. + src = _load_model_source() + pending = src.index("_pending_load_kwargs") + clamp = src.index("supports_kv_unified") + estimate = src.index("_estimate") + commit = src.index("_commit_effective_parallel_slots") + assert pending < clamp, "the requested count is captured before the clamp" + assert clamp < estimate, "the fit must be estimated from the effective slot count" + assert clamp < commit, "the committed effective count is the clamped one" + + +# ── Training-guard sizing ──────────────────────────────────────────── + + +def _write_swa_gguf(path: Path) -> str: + """Smallest DiffusionGemma-shaped header the KV estimator can size: the + canvas marker routing it to the diffusion runner, plus the sliding-window + dims that make llama.cpp's SWA cache slot-scaled.""" + + def _kv_str(key: str, value: str) -> bytes: + kb, vb = key.encode(), value.encode() + return ( + struct.pack(" bytes: + kb = key.encode() + return struct.pack(" float: + """Run the training guard over a local GGUF and return the size it budgeted.""" + import routes.inference as inf + + seen = {} + + core_training = _types.ModuleType("core.training") + core_training.get_training_backend = lambda: _types.SimpleNamespace( + is_training_active = lambda: True + ) + + def _can_load(**kwargs): + seen.update(kwargs) + return True, {"mode": "single_device"} + + training_vram = _types.ModuleType("routes.training_vram") + training_vram.can_load_chat_during_training = _can_load + monkeypatch.setitem(sys.modules, "core.training", core_training) + monkeypatch.setitem(sys.modules, "routes.training_vram", training_vram) + + monkeypatch.setattr(inf, "_classify_diffusion_gguf", lambda _config: diffusion) + monkeypatch.setattr(LlamaCppBackend, "_is_vulkan_backend", staticmethod(lambda *a, **k: False)) + monkeypatch.setattr(LlamaCppBackend, "_effective_gpu_count", staticmethod(lambda *a, **k: 1)) + monkeypatch.setattr(LlamaCppBackend, "_diffusion_gpu_arg", staticmethod(lambda *a, **k: "0")) + # Pin the --kv-unified probe so the estimate cannot depend on a locally + # installed llama-server. Default "no binary found" leaves the count alone. + monkeypatch.setattr( + LlamaCppBackend, + "probe_server_capabilities", + classmethod(lambda cls, binary = None: dict(caps or {})), + ) + + inf._guard_chat_load_against_training( + _types.SimpleNamespace(is_gguf = True, gguf_file = gguf_path, identifier = "local/model"), + model_identifier = "local/model", + hf_token = None, + load_in_4bit = False, + max_seq_length = 8192, + requested_gpu_ids = None, + n_parallel = n_parallel, + gpu_memory_mode = "auto", + ) + return seen["required_override_gb"] + + +def test_training_guard_sizes_a_diffusion_gguf_at_one_slot(monkeypatch, tmp_path): + # Diffusion ignores --parallel, so slots must not inflate the estimate and 409 + # a load that would have fitted beside training. + gguf = _write_swa_gguf(tmp_path / "diffusion.gguf") + one = _guard_required_gb(monkeypatch, gguf, n_parallel = 1, diffusion = True) + many = _guard_required_gb(monkeypatch, gguf, n_parallel = 8, diffusion = True) + assert one == many + + +def test_training_guard_still_sizes_slots_for_an_ordinary_gguf(monkeypatch, tmp_path): + # llama-server does allocate per-slot SWA cells, so the reduction above must + # be scoped to diffusion and not flatten every GGUF to one slot. + gguf = _write_swa_gguf(tmp_path / "chat.gguf") + one = _guard_required_gb(monkeypatch, gguf, n_parallel = 1, diffusion = False) + many = _guard_required_gb(monkeypatch, gguf, n_parallel = 8, diffusion = False) + assert many > one + + +def test_training_guard_sizes_one_slot_when_the_binary_has_no_kv_unified(monkeypatch, tmp_path): + # load_model clamps a multi-slot request to 1 on such a build, where each slot + # carries its own SWA stream, so sizing the asked count would 409 a load that fits. + gguf = _write_swa_gguf(tmp_path / "chat.gguf") + old = {"found": True, "supports_kv_unified": False} + one = _guard_required_gb(monkeypatch, gguf, n_parallel = 1, diffusion = False, caps = old) + many = _guard_required_gb(monkeypatch, gguf, n_parallel = 8, diffusion = False, caps = old) + assert one == many + + +def test_training_guard_sizes_every_slot_when_kv_unified_exists(monkeypatch, tmp_path): + # The clamp is scoped to binaries that cannot serve the slots; a capable one + # really does allocate the SWA window per slot. + gguf = _write_swa_gguf(tmp_path / "chat.gguf") + new = {"found": True, "supports_kv_unified": True} + one = _guard_required_gb(monkeypatch, gguf, n_parallel = 1, diffusion = False, caps = new) + many = _guard_required_gb(monkeypatch, gguf, n_parallel = 8, diffusion = False, caps = new) + assert many > one + + +def test_training_guard_keeps_slots_for_an_unclassified_gguf(monkeypatch, tmp_path): + # None = inconclusive header, so keep the larger estimate rather than + # under-size against training. + gguf = _write_swa_gguf(tmp_path / "unknown.gguf") + one = _guard_required_gb(monkeypatch, gguf, n_parallel = 1, diffusion = None) + many = _guard_required_gb(monkeypatch, gguf, n_parallel = 8, diffusion = None) + assert many > one diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts index 08d17f2a65..5f6c6cc589 100644 --- a/studio/frontend/src/features/chat/api/chat-adapter.ts +++ b/studio/frontend/src/features/chat/api/chat-adapter.ts @@ -1604,6 +1604,7 @@ async function autoLoadSmallestModel(): Promise<{ ? { gpu_ids: effectiveGpuIds ?? undefined, gpu_memory_mode: effectiveGpuMemoryMode, + n_parallel: config.nParallel ?? null, } : {}), })) @@ -1637,6 +1638,8 @@ async function autoLoadSmallestModel(): Promise<{ gpu_layers: effectiveGpuLayers, n_cpu_moe: effectiveNCpuMoe, gpu_ids: effectiveGpuIds ?? undefined, + // Per-model too, or the auto-load reverts a remembered override. + n_parallel: config.nParallel ?? null, } : {}), }); @@ -1689,6 +1692,11 @@ async function autoLoadSmallestModel(): Promise<{ effectiveGpuLayers, config.customContextLength ?? null, ); + // Slots this auto-load committed. Diffusion ignores --parallel, so a count + // there would mint a phantom override a saved preset carries onto a GGUF. + const committedSlots = (loadResp.is_diffusion ?? false) + ? null + : (config.nParallel ?? null); useChatRuntimeStore.setState({ ggufContextLength: loadResp.context_length ?? 131072, ggufMaxContextLength: @@ -1703,6 +1711,9 @@ async function autoLoadSmallestModel(): Promise<{ ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false), kvCacheDtype: loadResp.cache_type_kv ?? null, loadedKvCacheDtype: loadResp.cache_type_kv ?? null, + // Click-time value, not the resolved backend echo (see performLoad). + nParallel: committedSlots, + loadedNParallel: committedSlots, tensorParallel: loadResp.tensor_parallel ?? false, loadedTensorParallel: loadResp.tensor_parallel ?? false, ...loadedGpuMemoryFields(loadResp), @@ -1728,6 +1739,10 @@ async function autoLoadSmallestModel(): Promise<{ ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false), kvCacheDtype: loadResp.cache_type_kv ?? null, loadedKvCacheDtype: loadResp.cache_type_kv ?? null, + // GGUF-only and never sent here: a staged override would be saved for + // a model that cannot use it. + nParallel: null, + loadedNParallel: null, tensorParallel: loadResp.tensor_parallel ?? false, loadedTensorParallel: loadResp.tensor_parallel ?? false, // Non-GGUF response: clears any stale GPU baseline a prior manual-GPU @@ -2001,6 +2016,10 @@ async function autoLoadSmallestModel(): Promise<{ ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false), kvCacheDtype: loadResp.cache_type_kv ?? null, loadedKvCacheDtype: loadResp.cache_type_kv ?? null, + // The request above omits n_parallel: a staged override left from a + // preset would read as applied and be re-sent by the next Apply. + nParallel: null, + loadedNParallel: null, tensorParallel: loadResp.tensor_parallel ?? false, loadedTensorParallel: loadResp.tensor_parallel ?? false, ...loadedGpuMemoryFields(loadResp), diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts index 60b737fb68..8ad2691391 100644 --- a/studio/frontend/src/features/chat/api/chat-api.ts +++ b/studio/frontend/src/features/chat/api/chat-api.ts @@ -192,6 +192,8 @@ export async function validateModel( // --fit, while a pinned layer count is owned by the user. Tell validate // so it applies the same training-guard policy as /load. gpu_memory_mode: payload.gpu_memory_mode, + // Slots scale the KV estimate; keep validate sized like the load. + n_parallel: payload.n_parallel, }), }); return parseJsonOrThrow(response); diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx index 7b310c50d4..6070bd2e40 100644 --- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx +++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx @@ -397,6 +397,7 @@ export function ChatSettingsPanel({ const nCpuMoe = useChatRuntimeStore((s) => s.nCpuMoe); const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel); const specDraftNMax = useChatRuntimeStore((s) => s.specDraftNMax); + const nParallel = useChatRuntimeStore((s) => s.nParallel); const speculativeType = useChatRuntimeStore((s) => s.speculativeType); const specFallbackReason = useChatRuntimeStore((s) => s.specFallbackReason); const mtpUpdatable = @@ -504,6 +505,7 @@ export function ChatSettingsPanel({ tensorParallel, speculativeType, specDraftNMax, + nParallel, params.maxSeqLength, ]); const activePresetLoadSummary = useMemo( @@ -522,6 +524,7 @@ export function ChatSettingsPanel({ tensorParallel, speculativeType, specDraftNMax, + nParallel, params.maxSeqLength, ], ); diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts index bc7227e70d..5f0149d909 100644 --- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts +++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts @@ -567,6 +567,8 @@ export function useChatModelRuntime() { applyActiveModelStatusToStore(residentStatus, { previousCheckpoint: selectedCheckpoint, previousGgufVariant, + // Id and variant matched above: same model, only the tab moved. + readoptingSameModel: true, }); syncModelCapabilities(modelId, residentStatus); return; @@ -669,6 +671,14 @@ export function useChatModelRuntime() { let previousWasUnloaded = false; const pendingLoadConfig = typeof selection !== "string" ? selection.config : undefined; + // The outgoing model's slot INTENT (blank = follow the server + // default), which the resolved baseline cannot express. previousConfig + // is the snapshot the picker took before pre-applying the target's + // config, so the live control is only the outgoing one without it. + const previousNParallel = + typeof selection !== "string" && selection.previousConfig + ? (selection.previousConfig.nParallel ?? null) + : useChatRuntimeStore.getState().nParallel; if (pendingLoadConfig) { applyPerModelConfigToRuntime(pendingLoadConfig); } @@ -761,6 +771,8 @@ export function useChatModelRuntime() { : stateBeforeUnload.speculativeType; let loadSpecDraftNMax = pendingLoadConfig?.specDraftNMax ?? stateBeforeUnload.specDraftNMax; + let loadNParallel = + pendingLoadConfig?.nParallel ?? stateBeforeUnload.nParallel; try { // Lightweight pre-flight validation: avoid unloading a working model // if the new identifier is clearly invalid (e.g. bad HF id / path). @@ -792,6 +804,10 @@ export function useChatModelRuntime() { const validateGpuLayers = resetsPerModelSettings ? GPU_LAYERS_AUTO : loadGpuLayers; + // Per-model: the reset re-baselines to the staged config, like the load. + const validateNParallel = resetsPerModelSettings + ? (pendingLoadConfig?.nParallel ?? null) + : loadNParallel; const validateMaxSeqLength = resolveFitMaxSeqLength( isGguf, loadGpuMemoryMode, @@ -820,7 +836,12 @@ export function useChatModelRuntime() { cache_type_kv: loadKvCacheDtype, tensor_parallel: loadTensorParallel, gpu_ids: validateGpuIds ?? undefined, - ...(isGguf ? { gpu_memory_mode: loadGpuMemoryMode } : {}), + ...(isGguf + ? { + gpu_memory_mode: loadGpuMemoryMode, + n_parallel: validateNParallel, + } + : {}), }); // Upgrade consent runs before the security dialogs; Accept installs and the load continues. if (validation.requires_transformers_upgrade) { @@ -903,6 +924,10 @@ export function useChatModelRuntime() { loadedSpeculativeType: persistedSpeculativeType, specDraftNMax: null, loadedSpecDraftNMax: null, + // Per-model too: a different model follows the server default + // unless its staged config overrides it. + nParallel: null, + loadedNParallel: null, // Per-model GPU knobs must not follow onto a different model // (gpuMemoryMode is a standing preference and is kept). selectedGpuIds: null, @@ -918,6 +943,7 @@ export function useChatModelRuntime() { ? normalizeSpeculativeType(pendingLoadConfig.speculativeType) : persistedSpeculativeType; loadSpecDraftNMax = pendingLoadConfig?.specDraftNMax ?? null; + loadNParallel = pendingLoadConfig?.nParallel ?? null; // Keep the click-time snapshot in lock-step with the store reset so // the load below sizes against the cleared per-model knobs, not the // previous model's (gpuMemoryMode is standing, so left as captured). @@ -984,6 +1010,8 @@ export function useChatModelRuntime() { cache_type_kv: loadKvCacheDtype, speculative_type: loadSpeculativeType, spec_draft_n_max: loadSpecDraftNMax, + // GGUF-only: slots mean nothing for a transformers load. + n_parallel: isGguf ? loadNParallel : null, tensor_parallel: loadTensorParallel, gpu_memory_mode: loadGpuMemoryMode, gpu_layers: loadGpuLayers, @@ -1034,6 +1062,14 @@ export function useChatModelRuntime() { const loadedSpec = normalizeSpeculativeType( loadResponse.speculative_type, ); + // Slots the load actually committed. Non-GGUF never sends them and + // diffusion ignores --parallel, so a click-time count on either + // would mint a phantom override a saved preset carries onto a GGUF. + const committedSlots = + (loadResponse.is_gguf ?? false) && + !(loadResponse.is_diffusion ?? false) + ? (loadNParallel ?? null) + : null; const nativeCtx = loadResponse.is_gguf ? (loadResponse.context_length ?? 131072) : null; @@ -1109,6 +1145,10 @@ export function useChatModelRuntime() { loadedSpeculativeType: loadedSpec, specDraftNMax: loadResponse.spec_draft_n_max ?? null, loadedSpecDraftNMax: loadResponse.spec_draft_n_max ?? null, + // Keep the click-time value: the echo is the resolved count, and + // adopting it would pin a blank "server default" control. + nParallel: committedSlots, + loadedNParallel: committedSlots, customContextLength: keepCustomCtx, loadedCustomContextLength: keepCustomCtx, defaultChatTemplate: loadResponse.chat_template ?? null, @@ -1211,6 +1251,7 @@ export function useChatModelRuntime() { stateBeforeUnload.loadedSpeculativeType, spec_draft_n_max: stateBeforeUnload.loadedSpecDraftNMax, + n_parallel: stateBeforeUnload.loadedNParallel, // Restore the previous model in the split mode it was running, // not the default layer split. tensor_parallel: stateBeforeUnload.loadedTensorParallel ?? false, @@ -1237,6 +1278,9 @@ export function useChatModelRuntime() { // model's; the loaded baselines below come from its reload echo. speculativeType: stateBeforeUnload.loadedSpeculativeType ?? null, specDraftNMax: stateBeforeUnload.loadedSpecDraftNMax ?? null, + // Control keeps its intent; only the baseline takes the echo. + nParallel: previousNParallel, + loadedNParallel: stateBeforeUnload.loadedNParallel ?? null, loadedSpeculativeType: rollbackSpeculativeType, loadedSpecDraftNMax: rollbackResponse.spec_draft_n_max ?? null, diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts index f85ff3246b..47d40009c8 100644 --- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts +++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts @@ -1,6 +1,9 @@ // SPDX-License-Identifier: AGPL-3.0-only // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 +// Barrel import (lint rule); the model-picker cycle is fine because the call +// happens at runtime, not module eval. +import { resolveInitialConfig } from "@/features/model-picker"; import { getInferenceStatus } from "../api/chat-api"; import { mergeBackendRecommendedInference, @@ -131,6 +134,9 @@ export type ApplyInferenceStatusOptions = { * status -- without it a variant-only switch underneath the tab reads as * steady state and the hydration reseed keeps the old quant's baselines. */ previousGgufVariant?: string | null; + /** The caller verified the status is the model this tab just picked, so the + * slot control it holds belongs to that model and must survive. */ + readoptingSameModel?: boolean; }; /** Mirror refresh() hydration so adopted CLI models get reasoning/tools flags. */ @@ -201,6 +207,22 @@ export function applyActiveModelStatusToStore( // While a load is in flight, performLoad owns the load params. Seeding them // from a stale poll here would clobber the values the load dialog just set. const seedLoadParams = !prevState.modelLoading; + // A model/variant change underneath this tab, as opposed to re-adopting the + // model the tab just picked, where hydratingExistingModel fires on the stale + // checkpoint. The echo cannot stand in: a new model can report the old count. + const slotsModelChanged = + hydratingExistingModel && !options.readoptingSameModel; + // This model's remembered override, read only on a fresh store or a model + // change, so a steady poll cannot re-pin a control the user just blanked. + const slotsUnseeded = + prevState.loadedNParallel === null && prevState.nParallel === null; + const remembered = + status.is_gguf && (slotsUnseeded || slotsModelChanged) + ? resolveInitialConfig(checkpointId, status.gguf_variant ?? null) + : null; + const rememberedNParallel = remembered?.remembered + ? (remembered.config.nParallel ?? null) + : null; // A Manual + Auto-layers load sent its positive context pin as max_seq_length, // and status only exposes the RESOLVED context; re-seed the pin from the // requested value (parity with the load paths' keepCustomCtx). Baselines @@ -322,6 +344,35 @@ export function applyActiveModelStatusToStore( tensorParallel: status.tensor_parallel, loadedTensorParallel: status.tensor_parallel, }), + // Baseline only, never the control: the echo is the RESOLVED count and would + // pin a blank "server default" control. The rollback re-sends the baseline, + // so without this a rollback after a tab reload loses the override. + ...(seedLoadParams && + status.requested_parallel_slots != null && + (prevState.loadedNParallel === null || hydratingExistingModel) && { + loadedNParallel: status.requested_parallel_slots, + }), + // A slotless model must not keep the previous GGUF's baseline: the rollback + // re-sends it. /status omits the echo for non-GGUF and sends an explicit + // null for diffusion, so an absent field on a GGUF is an older backend. + ...(seedLoadParams && + (status.is_gguf === false || status.requested_parallel_slots === null) && { + loadedNParallel: null, + }), + // Per-model: a change underneath this tab blanks the control like + // performLoad's cross-model reset, or the old count follows onto the new + // model. The baseline above still carries the rollback. + ...(seedLoadParams && slotsModelChanged && { nParallel: null }), + // AFTER that clear, which both a first hydration and a model change trip: + // either would leave the control blank while the model runs on a remembered + // override, so the next Apply would save the blank over it. Adopted only + // when the running count matches, proving it is this model's own. + ...(seedLoadParams && + (slotsUnseeded || slotsModelChanged) && + rememberedNParallel != null && + rememberedNParallel === status.requested_parallel_slots && { + nParallel: rememberedNParallel, + }), // Re-seed on first hydration, model/variant changes, or a same-model backend // placement change. gpuStatusFields preserves dirty local edits in the last // case while advancing their loaded baselines. diff --git a/studio/frontend/src/features/chat/presets/preset-load-config.ts b/studio/frontend/src/features/chat/presets/preset-load-config.ts index 1083655cf2..c0a65c7886 100644 --- a/studio/frontend/src/features/chat/presets/preset-load-config.ts +++ b/studio/frontend/src/features/chat/presets/preset-load-config.ts @@ -12,6 +12,8 @@ import { DEFAULT_MAX_SEQ_LENGTH, KV_CACHE_DTYPES, MTP_SPECULATIVE_TYPES, + N_PARALLEL_MAX, + N_PARALLEL_MIN, SPECULATIVE_TYPES, normalizeMaxSeqLength, type PerModelConfig, @@ -30,6 +32,7 @@ export type PresetLoadConfig = Pick< | "kvCacheDtype" | "speculativeType" | "specDraftNMax" + | "nParallel" | "tensorParallel" | "gpuMemoryMode" | "gpuLayers" @@ -45,6 +48,7 @@ export const EMPTY_PRESET_LOAD_CONFIG: PresetLoadConfig = { kvCacheDtype: null, speculativeType: null, specDraftNMax: null, + nParallel: null, tensorParallel: false, }; @@ -107,6 +111,14 @@ export function normalizePresetLoadConfig( ? speculativeType : null, specDraftNMax, + nParallel: + typeof partial.nParallel === "number" && + Number.isFinite(partial.nParallel) + ? Math.max( + N_PARALLEL_MIN, + Math.min(N_PARALLEL_MAX, Math.round(partial.nParallel)), + ) + : null, tensorParallel: typeof partial.tensorParallel === "boolean" ? partial.tensorParallel @@ -151,6 +163,7 @@ export function capturePresetLoadConfig(): PresetLoadConfig | undefined { kvCacheDtype: snapshot.kvCacheDtype ?? null, speculativeType: normalizeSpeculativeType(snapshot.speculativeType), specDraftNMax: snapshot.specDraftNMax ?? null, + nParallel: snapshot.nParallel ?? null, tensorParallel: snapshot.tensorParallel ?? false, ...(snapshot.gpuMemoryMode === "manual" ? { gpuMemoryMode: "manual" as const } @@ -206,6 +219,7 @@ export function applyPresetLoadConfig( kvCacheDtype: config.kvCacheDtype ?? null, speculativeType: config.speculativeType ?? null, specDraftNMax: config.specDraftNMax ?? null, + nParallel: config.nParallel ?? null, tensorParallel: config.tensorParallel ?? false, chatTemplateOverride: null, gpuMemoryMode: config.gpuMemoryMode, @@ -231,6 +245,9 @@ export function formatPresetLoadConfigSummary( if (config.speculativeType && config.speculativeType !== "auto") { parts.push(`Spec ${config.speculativeType}`); } + if (config.nParallel != null) { + parts.push(`${config.nParallel} slots`); + } if (config.gpuMemoryMode === "manual") { parts.push("GPU manual"); } diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx index 44436b92df..890dd022a0 100644 --- a/studio/frontend/src/features/chat/shared-composer.tsx +++ b/studio/frontend/src/features/chat/shared-composer.tsx @@ -1130,6 +1130,8 @@ export function SharedComposer({ ? { gpu_ids: effectiveSelectedGpuIds ?? undefined, gpu_memory_mode: effectiveGpuMemoryMode, + // Slots scale the KV estimate; keep validate sized like the load. + n_parallel: ownConfig.nParallel ?? null, } : {}), }); @@ -1198,6 +1200,7 @@ export function SharedComposer({ n_cpu_moe: effectiveNCpuMoe, tensor_split: compareLoadKnobs.splitRatio ?? undefined, gpu_ids: effectiveSelectedGpuIds ?? undefined, + n_parallel: ownConfig.nParallel ?? null, } : {}), }); @@ -1229,6 +1232,12 @@ export function SharedComposer({ effectiveCustomContextLength, ) : null; + // Slots this compare load committed. Diffusion ignores --parallel, so a + // count there would mint a phantom override a preset carries onto a GGUF. + const committedSlots = + targetIsGguf && !(resp.is_diffusion ?? false) + ? (ownConfig.nParallel ?? null) + : null; useChatRuntimeStore.setState({ supportsReasoning: resp.supports_reasoning ?? false, reasoningAlwaysOn: resp.reasoning_always_on ?? false, @@ -1237,6 +1246,9 @@ export function SharedComposer({ supportsTools: resp.supports_tools ?? false, kvCacheDtype: resp.cache_type_kv ?? null, loadedKvCacheDtype: resp.cache_type_kv ?? null, + // Click-time value, not the resolved echo (see the single-model load). + nParallel: committedSlots, + loadedNParallel: committedSlots, tensorParallel: resp.tensor_parallel ?? false, loadedTensorParallel: resp.tensor_parallel ?? false, defaultChatTemplate: resp.chat_template ?? null, diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts index 98b8676c10..2984611780 100644 --- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts +++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts @@ -968,6 +968,12 @@ type ChatRuntimeStore = { /** User --spec-draft-n-max override (null = platform default). */ specDraftNMax: number | null; loadedSpecDraftNMax: number | null; + /** User --parallel slots override for GGUF loads (null = server default). + * Never re-seeded from an echo: the resolved count would pin a blank control. */ + nParallel: number | null; + /** Slots the last successful load sent (null = default); the rollback + * re-sends it so a failed switch can't lose the override. */ + loadedNParallel: number | null; /** Tensor-parallel split (--split-mode tensor) toggle, GGUF multi-GPU only. */ tensorParallel: boolean; /** Backend-reported tensor-parallel state; null until first hydrated. */ @@ -1491,6 +1497,8 @@ export const useChatRuntimeStore = create((set, get) => ({ specFallbackReason: null, specDraftNMax: null, loadedSpecDraftNMax: null, + nParallel: null, + loadedNParallel: null, tensorParallel: false, loadedTensorParallel: null, gpuMemoryMode: readPersistedGpuMemoryMode(), @@ -1874,6 +1882,8 @@ export const useChatRuntimeStore = create((set, get) => ({ specFallbackReason: null, specDraftNMax: null, loadedSpecDraftNMax: null, + nParallel: null, + loadedNParallel: null, tensorParallel: false, loadedTensorParallel: null, // Standing preference: survives unload, unlike the per-model knobs above. diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts index 3681b0f0cb..b67a9eda26 100644 --- a/studio/frontend/src/features/chat/types/api.ts +++ b/studio/frontend/src/features/chat/types/api.ts @@ -65,6 +65,11 @@ export interface LoadModelRequest { * when speculative_type resolves to "mtp" or "mtp+ngram". */ spec_draft_n_max?: number | null; + /** + * Parallel decode slots for llama-server (--parallel), 1..64. Omit/null = + * the launch default. The VRAM fitter may launch fewer to stay on GPU. + */ + n_parallel?: number | null; /** * Split the model across GPUs by tensor (--split-mode tensor) instead * of by layer for GGUF models. Multi-GPU only; no effect on a single GPU. @@ -202,6 +207,12 @@ export interface LoadModelResponse { gpu_ids?: number[] | null; /** User-requested GPU placement pool before fit-time narrowing. */ requested_gpu_ids?: number[] | null; + /** Slots the load was invoked with (else the --parallel default). Null for + * non-GGUF loads. */ + requested_parallel_slots?: number | null; + /** Slots llama-server actually runs, after any fit-time reduction. Null for + * non-GGUF loads. */ + parallel_slots?: number | null; } export interface UnloadModelRequest { @@ -263,6 +274,12 @@ export interface InferenceStatusResponse { gpu_ids?: number[] | null; /** User-requested GPU placement pool before fit-time narrowing. */ requested_gpu_ids?: number[] | null; + /** Slots the active load was invoked with (else the --parallel default). + * Null when no GGUF model is loaded. */ + requested_parallel_slots?: number | null; + /** Slots llama-server actually runs, after any fit-time reduction. Null when + * no GGUF model is loaded. */ + parallel_slots?: number | null; n_layers?: number | null; /** Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not MoE. */ n_moe_layers?: number; diff --git a/studio/frontend/src/features/model-picker/components/model-config-page.tsx b/studio/frontend/src/features/model-picker/components/model-config-page.tsx index 90202a2bcf..a753e9016e 100644 --- a/studio/frontend/src/features/model-picker/components/model-config-page.tsx +++ b/studio/frontend/src/features/model-picker/components/model-config-page.tsx @@ -46,6 +46,8 @@ import { MAX_SEQ_LENGTH_MIN, MAX_SEQ_LENGTH_STEP, MTP_SPECULATIVE_TYPES, + N_PARALLEL_MAX, + N_PARALLEL_MIN, type PerModelConfig, SPECULATIVE_TYPES, deletePerModelConfig, @@ -87,6 +89,7 @@ function hasNonDefaultAdvanced(config: PerModelConfig): boolean { config.kvCacheDtype != null || (config.speculativeType ?? "auto") !== "auto" || config.specDraftNMax != null || + config.nParallel != null || config.tensorParallel || config.chatTemplateOverride != null || (config.gpuMemoryMode ?? "auto") !== "auto" || @@ -541,6 +544,44 @@ function GgufAdvancedSettings({ )} +
+
+ Parallel Slots + + llama-server decode slots (--parallel) for concurrent requests. + Leave blank for the server default. More slots share the context + pool and use more VRAM; if they don't fit on GPU, fewer slots are + launched. + +
+ { + const raw = event.target.value; + if (raw === "") { + update({ nParallel: null }); + return; + } + const parsed = Number.parseInt(raw, 10); + if (Number.isFinite(parsed)) { + update({ + nParallel: Math.max( + N_PARALLEL_MIN, + Math.min(N_PARALLEL_MAX, parsed), + ), + }); + } + }} + aria-label="Parallel decode slots" + className={NUMBER_INPUT_CLASS} + /> +
+
Tensor Parallelism diff --git a/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx b/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx index 2d12c503a4..0d5f0fd663 100644 --- a/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx +++ b/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx @@ -43,6 +43,7 @@ function configSignature(config: PerModelConfig): string { config.kvCacheDtype ?? "", config.speculativeType ?? "", config.specDraftNMax ?? "", + config.nParallel ?? "", config.tensorParallel ? "1" : "0", config.chatTemplateOverride == null ? "" diff --git a/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts b/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts index 9d09ee6897..b0a6411019 100644 --- a/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts +++ b/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts @@ -20,6 +20,7 @@ export function useActiveModelConfig(): ActiveModelConfigState { const kvCacheDtype = useChatRuntimeStore((s) => s.kvCacheDtype); const speculativeType = useChatRuntimeStore((s) => s.speculativeType); const specDraftNMax = useChatRuntimeStore((s) => s.specDraftNMax); + const nParallel = useChatRuntimeStore((s) => s.nParallel); const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel); const chatTemplateOverride = useChatRuntimeStore( (s) => s.chatTemplateOverride, @@ -44,6 +45,7 @@ export function useActiveModelConfig(): ActiveModelConfigState { kvCacheDtype: kvCacheDtype ?? null, speculativeType: speculativeType ?? "auto", specDraftNMax: specDraftNMax ?? null, + nParallel: nParallel ?? null, tensorParallel: tensorParallel ?? false, chatTemplateOverride: chatTemplateOverride ?? null, }; @@ -65,6 +67,7 @@ export function useActiveModelConfig(): ActiveModelConfigState { kvCacheDtype, speculativeType, specDraftNMax, + nParallel, tensorParallel, chatTemplateOverride, gpuMemoryMode, diff --git a/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts b/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts index c21d3e164a..829c522cb2 100644 --- a/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts +++ b/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts @@ -39,6 +39,7 @@ export function applyPerModelConfigToRuntime(config: PerModelConfig): void { normalizeSpeculativeType(config.speculativeType) ?? readPersistedSpeculativeType(), specDraftNMax: config.specDraftNMax ?? null, + nParallel: config.nParallel ?? null, tensorParallel: config.tensorParallel ?? false, chatTemplateOverride: cleanTemplate(config.chatTemplateOverride), // GPU Memory knobs are per-model (GGUF-only). Absent = defaults; the mode is @@ -77,6 +78,7 @@ export function currentRuntimePerModelConfig( kvCacheDtype: s.kvCacheDtype ?? null, speculativeType: normalizeSpeculativeType(s.speculativeType), specDraftNMax: s.specDraftNMax ?? null, + nParallel: s.nParallel ?? null, tensorParallel: s.tensorParallel ?? false, chatTemplateOverride: cleanTemplate(s.chatTemplateOverride), // Snapshot the live GPU knobs too so a failed switch rolls the previous @@ -101,6 +103,7 @@ export function perModelConfigsEqual( normalizeSpeculativeType(a.speculativeType) === normalizeSpeculativeType(b.speculativeType) && (a.specDraftNMax ?? null) === (b.specDraftNMax ?? null) && + (a.nParallel ?? null) === (b.nParallel ?? null) && Boolean(a.tensorParallel) === Boolean(b.tensorParallel) && cleanTemplate(a.chatTemplateOverride) === cleanTemplate(b.chatTemplateOverride) && diff --git a/studio/frontend/src/features/model-picker/model-config/per-model-config.ts b/studio/frontend/src/features/model-picker/model-config/per-model-config.ts index ba6d4cec99..196ac9e5a1 100644 --- a/studio/frontend/src/features/model-picker/model-config/per-model-config.ts +++ b/studio/frontend/src/features/model-picker/model-config/per-model-config.ts @@ -15,6 +15,7 @@ export interface PerModelConfig { kvCacheDtype: string | null; speculativeType: string | null; specDraftNMax: number | null; + nParallel: number | null; tensorParallel: boolean; chatTemplateOverride: string | null; // GPU Memory controls (per-model, GGUF-only), optional so older blobs still @@ -33,10 +34,16 @@ export const DEFAULT_PER_MODEL_CONFIG: PerModelConfig = { kvCacheDtype: null, speculativeType: null, specDraftNMax: null, + nParallel: null, tensorParallel: false, chatTemplateOverride: null, }; +// Mirrors llama_server_args.py PARALLEL_MIN/MAX (LoadRequest.n_parallel +// bounds). null = follow the server-wide default. +export const N_PARALLEL_MIN = 1; +export const N_PARALLEL_MAX = 64; + export const MAX_SEQ_LENGTH_MIN = 128; export const MAX_SEQ_LENGTH_MAX = 1048576; export const MAX_SEQ_LENGTH_STEP = 128; @@ -92,6 +99,7 @@ const STORED_CONFIG_FIELDS = new Set([ "kvCacheDtype", "speculativeType", "specDraftNMax", + "nParallel", "tensorParallel", "chatTemplateOverride", "gpuMemoryMode", @@ -292,6 +300,8 @@ function legacyEntryToConfig(raw: Record): PerModelConfig { typeof raw.speculativeType === "string" ? raw.speculativeType : null, specDraftNMax: typeof raw.specDraftNMax === "number" ? raw.specDraftNMax : null, + // Legacy blobs predate the parallel-slots knob. + nParallel: null, tensorParallel: typeof raw.tensorParallel === "boolean" ? raw.tensorParallel : false, chatTemplateOverride: null, @@ -459,6 +469,10 @@ function normalizeV1(partial: RawConfig): PerModelConfig { : null, speculativeType, specDraftNMax, + nParallel: + typeof partial.nParallel === "number" && Number.isFinite(partial.nParallel) + ? Math.max(N_PARALLEL_MIN, Math.min(N_PARALLEL_MAX, Math.round(partial.nParallel))) + : null, tensorParallel: typeof partial.tensorParallel === "boolean" ? partial.tensorParallel @@ -597,6 +611,7 @@ export function isDefaultConfig(config: PerModelConfig): boolean { (config.kvCacheDtype ?? null) === DEFAULT_PER_MODEL_CONFIG.kvCacheDtype && config.speculativeType === DEFAULT_PER_MODEL_CONFIG.speculativeType && config.specDraftNMax == null && + config.nParallel == null && Boolean(config.tensorParallel) === Boolean(DEFAULT_PER_MODEL_CONFIG.tensorParallel) && (config.chatTemplateOverride ?? null) === null && diff --git a/tests/studio/test_chat_preset_load_config.py b/tests/studio/test_chat_preset_load_config.py index 1588c7d96d..6234ab395c 100644 --- a/tests/studio/test_chat_preset_load_config.py +++ b/tests/studio/test_chat_preset_load_config.py @@ -68,3 +68,18 @@ def test_backend_chat_preset_accepts_load_config(): routes = _read("studio/backend/routes/chat_history.py") assert "class ChatPresetLoadConfig" in routes assert "loadConfig: Optional[ChatPresetLoadConfig]" in routes + + +def test_preset_load_config_carries_parallel_slots(): + # Captured, clamped on read, applied, and accepted by the extra="forbid" + # backend model (a missing backend field would 422 every settings sync). + source = _read("studio/frontend/src/features/chat/presets/preset-load-config.ts") + assert '| "nParallel"' in source + assert "nParallel: snapshot.nParallel ?? null" in source + assert "nParallel: config.nParallel ?? null" in source + assert "N_PARALLEL_MAX, Math.round(partial.nParallel)" in source + routes = _read("studio/backend/routes/chat_history.py") + assert ( + "nParallel: Optional[int] = Field(default = None, ge = PARALLEL_MIN, le = PARALLEL_MAX)" + in routes + ) diff --git a/tests/studio/test_model_picker_contracts.py b/tests/studio/test_model_picker_contracts.py index 00ee83efc7..d9a8efc9a4 100644 --- a/tests/studio/test_model_picker_contracts.py +++ b/tests/studio/test_model_picker_contracts.py @@ -631,6 +631,253 @@ def test_legacy_migration_is_idempotent_and_non_destructive(): assert "if (isDefaultConfig(migrated) || Object.hasOwn(map, key)) {" in src +def test_parallel_slots_setting_wired_end_to_end(): + """The per-load Parallel Slots knob (llama-server --parallel) must flow from + the run-settings form through persistence, every /load builder, the validate + preflight and the cross-model reset; a lost hop silently reverts the model to + the server-wide slot default.""" + config = _read("features/model-picker/model-config/per-model-config.ts") + # Persisted per model, clamped on every read/write, and null (= server + # default) counts as default so blank configs are not stored. + assert '"nParallel",' in config + assert "N_PARALLEL_MAX, Math.round(partial.nParallel)" in config + assert "config.nParallel == null &&" in config + page = _read("features/model-picker/components/model-config-page.tsx") + # Rendered in the GGUF advanced section, which a remembered override reopens. + assert "Parallel Slots" in page + assert "config.nParallel != null ||" in page + assert 'aria-label="Parallel decode slots"' in page + api_types = _read("features/chat/types/api.ts") + assert "n_parallel?: number | null;" in api_types + runtime = _read("features/chat/hooks/use-chat-model-runtime.ts") + # Click-time snapshot, /load body, validate preflight, cross-model reset and + # failed-switch rollback all carry the value. + assert "pendingLoadConfig?.nParallel" in runtime + # GGUF-gated, like the compare pane: a transformers load has no slots. + assert "n_parallel: isGguf ? loadNParallel : null," in runtime + assert "n_parallel: validateNParallel," in runtime + assert "loadNParallel = pendingLoadConfig?.nParallel ?? null;" in runtime + assert "n_parallel: stateBeforeUnload.loadedNParallel," in runtime + chat_api = _read("features/chat/api/chat-api.ts") + assert "n_parallel: payload.n_parallel," in chat_api + composer = _read("features/chat/shared-composer.tsx") + # The compare pane is a second /load builder; its preflight sizes like its load. + assert composer.count("n_parallel: ownConfig.nParallel ?? null,") == 2 + adapter = _read("features/chat/api/chat-adapter.ts") + # The startup auto-load is a third builder reading the remembered config. + assert adapter.count("n_parallel: config.nParallel ?? null,") == 2 + # ... and records it as loaded through the diffusion-gated local below. + assert "loadedNParallel: committedSlots," in adapter + status = _read("features/chat/lib/apply-inference-status-to-store.ts") + # Hydration seeds the rollback BASELINE only; adopting the resolved echo into + # the control would pin a blank "server default" to a number. + assert "loadedNParallel: status.requested_parallel_slots," in status + assert "nParallel: status.requested_parallel_slots," not in status + sidebar = _read("features/model-picker/components/sidebar-model-config.tsx") + # The sidebar form remounts when an external change lands. + assert 'config.nParallel ?? "",' in sidebar + + +def test_parallel_slots_control_cleared_when_the_load_never_sent_them(): + """`nParallel` is the editable control ("blank = follow the server default") + and `loadedNParallel` the rollback baseline. A success path that sends no + slot count must blank the control, or a value staged for another model shows + as applied, is persisted into this model's config (`isDefaultConfig` keys on + nParallel) and is re-sent by the next Apply. Each assertion below is the only + thing pinning one such path.""" + status = " ".join(_read("features/chat/lib/apply-inference-status-to-store.ts").split()) + # A model/variant swap underneath this tab must reset the control like + # performLoad's cross-model reset, or model A's count follows onto model B. + # Narrowly gated -- see test_hydration_keeps_the_slot_control_when_readopting_the_running_model. + assert "...(seedLoadParams && slotsModelChanged && { nParallel: null })," in status + # ... while still never adopting the RESOLVED echo into the control. + assert "nParallel: status.requested_parallel_slots," not in status + + adapter = _read("features/chat/api/chat-adapter.ts") + # Slice the two success branches apart, bounding the second at the shared tail + # so it cannot swallow the fresh-default path below and stay green. + candidate = adapter.split("async function loadAutoLoadCandidate", 1)[1] + gguf_branch, non_gguf_rest = candidate.split('if (candidate.kind === "gguf") {', 1)[1].split( + "\n } else {\n", 1 + ) + non_gguf_branch = non_gguf_rest.split("if (!(loadResp.is_lora ?? false)) {", 1)[0] + # The cached-GGUF branch keeps the remembered override via the gated local... + assert "nParallel: committedSlots," in gguf_branch + assert "nParallel: null," not in gguf_branch + # ... the safetensors fallback sends no slots, so it clears both, or the count + # survives on a model whose form does not even render the field. + assert "nParallel: null," in non_gguf_branch + assert "loadedNParallel: null," in non_gguf_branch + + fresh_default = adapter.split("No downloaded models found. Fetching", 1)[1].split( + 'showAutoLoadSuccess("Loaded Qwen', 1 + )[0] + # The fresh-default download omits the slots, so its success state clears both, + # or the control reads as an unapplied edit against the seeded baseline. + assert "n_parallel" not in fresh_default.split("saveSpeculativeType", 1)[0] + assert "nParallel: null," in fresh_default + assert "loadedNParallel: null," in fresh_default + + +def test_hydration_clears_the_slot_baseline_for_a_slotless_model(): + """The baseline is what a rollback re-sends and what preset capture reads, so + a model that cannot have slots must not inherit the previous GGUF's count. + /status omits the echo for non-GGUF and sends an explicit null for diffusion; + an absent field on a GGUF is an older backend and must NOT wipe it.""" + src = _read("features/chat/lib/apply-inference-status-to-store.ts") + assert ( + "(status.is_gguf === false || status.requested_parallel_slots === null) && {" in src + ), "the slotless clear must key on is_gguf or an explicit null echo" + clear = src.index("status.is_gguf === false || status.requested_parallel_slots === null") + assert "loadedNParallel: null," in src[clear : clear + 200] + # Never `!= null`: that also matches the absent field an older backend sends. + assert "status.requested_parallel_slots !== null && {" not in src + + +def test_hydration_keeps_the_slot_control_when_readopting_the_running_model(): + """`hydratingExistingModel` is true whenever the incoming status disagrees + with what this tab last recorded, which includes RE-ADOPTING a model the tab + never lost: the resident-adopt branch restores the model's own per-model + config and only then hydrates, passing the EXTERNAL id as + `previousCheckpoint`. An ungated clear there wipes the slot count that branch + just restored, and the blank persists into `savePerModelConfig`, so a Save + the user reads as a no-op erases their remembered override. + + Only that branch knows the model is unchanged, so it says so explicitly. + Slot counts cannot stand in: the echo falls back to the server-wide default, + so a genuine A->B swap can echo exactly A's explicit count.""" + status = " ".join(_read("features/chat/lib/apply-inference-status-to-store.ts").split()) + assert ( + "const slotsModelChanged = hydratingExistingModel && !options.readoptingSameModel;" + in status + ) + assert "...(seedLoadParams && slotsModelChanged && { nParallel: null })," in status + # Never a slot-count proxy for "same model". + assert "prevState.loadedNParallel === (status.requested_parallel_slots" not in status + # The baseline seed stays ungated, or a rollback after a tab reload restores + # the model at the server default slots. + assert "loadedNParallel: status.requested_parallel_slots," in status + + runtime = " ".join(_read("features/chat/hooks/use-chat-model-runtime.ts").split()) + resident = runtime.split("if (!forceReload && isExternalModelId(selectedCheckpoint)) {", 1)[ + 1 + ].split("const stopDecision", 1)[0] + # What makes the scenario reachable: the branch restores the model's own + # config, then hydrates against the external id. + assert "applyPerModelConfigToRuntime(selection.previousConfig);" in resident + assert "previousCheckpoint: selectedCheckpoint," in resident + # Only reachable because the branch matched the id AND the variant first. + assert "resolveInferenceCheckpointId(residentStatus) === modelId" in resident + assert "readoptingSameModel: true," in resident + # The refresh() hydrate must NOT claim it: there the model really can change. + poll = runtime.split("setModels(listRes.models.map(toChatModelSummary));", 1)[1].split( + "} else if (!statusRes.active_model", 1 + )[0] + assert "applyActiveModelStatusToStore(statusRes, {" in poll + assert "readoptingSameModel" not in poll + + +def test_parallel_slots_are_never_recorded_for_a_diffusion_load(): + """A DiffusionGemma GGUF answers ``is_gguf: true``, but its runner ignores + ``--parallel``, so ``_parallel_slot_echo`` reports null slots for it. The + three load success paths must gate on ``is_diffusion`` too, or they record a + click-time count the load never committed. + + That phantom does not stay put: ``capturePresetLoadConfig`` snapshots + ``nParallel`` with no model gate and a preset carries no model identity, so + applying it over a TEXT GGUF sends the count as a real ``n_parallel``. + """ + runtime = " ".join(_read("features/chat/hooks/use-chat-model-runtime.ts").split()) + # One gated local feeds the control and the baseline, so they cannot drift. + assert "(loadResponse.is_gguf ?? false) && !(loadResponse.is_diffusion ?? false)" in runtime + assert "nParallel: committedSlots," in runtime + assert "loadedNParallel: committedSlots," in runtime + + adapter = " ".join(_read("features/chat/api/chat-adapter.ts").split()) + assert ( + "const committedSlots = (loadResp.is_diffusion ?? false) ? null " + ": (config.nParallel ?? null);" in adapter + ) + assert "nParallel: committedSlots," in adapter + assert "loadedNParallel: committedSlots," in adapter + + composer = " ".join(_read("features/chat/shared-composer.tsx").split()) + assert "targetIsGguf && !(resp.is_diffusion ?? false)" in composer + assert "nParallel: committedSlots," in composer + assert "loadedNParallel: committedSlots," in composer + + +def test_hydration_restores_a_remembered_slot_override(): + """The control is never seeded from the status echo, so a model running on a + remembered override shows a BLANK slot control after a browser reload or a + tab move to another GGUF. `ModelConfigPage.resolveInitial` prefers the live + store for the active model, so that blank is what the form edits: the next + Apply reloads at the server default and a Save writes the blank over the + remembered count. + + The seed is deliberately narrow: storage is read only on a fresh store or a + model change, never on a steady poll, and the value is adopted only when the + server already runs that exact count, which proves it is this model's own. + """ + src = _read("features/chat/lib/apply-inference-status-to-store.ts") + status = " ".join(src.split()) + assert ( + "resolveInitialConfig(checkpointId, status.gguf_variant ?? null)" in status + ), "the remembered override comes from per-model storage, not the echo" + assert ( + "const slotsUnseeded = prevState.loadedNParallel === null && " + "prevState.nParallel === null;" in status + ) + assert ( + "status.is_gguf && (slotsUnseeded || slotsModelChanged)" in status + ), "storage is read on a fresh store or a model change, never on a steady poll" + assert ( + "...(seedLoadParams && (slotsUnseeded || slotsModelChanged) &&" in status + ), "the seed fires in both cases the clear leaves the control blank" + assert ( + "rememberedNParallel != null && rememberedNParallel === " + "status.requested_parallel_slots && { nParallel: rememberedNParallel, }" in status + ) + # Both cases trip the model-change clear, so the seed only survives by + # being spread after it. + assert src.index("slotsModelChanged && { nParallel: null }") < src.index( + "nParallel: rememberedNParallel," + ) + + +def test_failed_switch_rollback_restores_the_slot_intent_not_the_resolved_count(): + """`loadedNParallel` holds a RESOLVED count even for a load that sent no + slots (the echo falls back to the server-wide default), so it is the right + value to re-send when recreating the previous server and the wrong one to put + back in the control: it turns "follow the server default" into an explicit + override that a later Save or preset capture pins. The outer catch only + repairs that for a staged config, so a plain string pick keeps the phantom. + + The intent comes from the picker's own pre-switch snapshot when there is one: + chat-page pre-applies the TARGET's config before calling selectModel, so the + live control describes the outgoing model only for a bare pick.""" + runtime = " ".join(_read("features/chat/hooks/use-chat-model-runtime.ts").split()) + assert ( + 'const previousNParallel = typeof selection !== "string" && ' + "selection.previousConfig ? (selection.previousConfig.nParallel ?? null) " + ": useChatRuntimeStore.getState().nParallel;" in runtime + ) + assert runtime.index("const previousNParallel") < runtime.index( + "applyPerModelConfigToRuntime(pendingLoadConfig);" + ), "a config staged on the selection must not replace it either" + picker = " ".join(_read("features/chat/chat-page.tsx").split()) + assert ( + "const previousConfig = currentRuntimePerModelConfig({ includeMaxSeqLength: true, }); " + "const hasAppliedConfig = applyModelLoadConfigToRuntime(" in picker + ), "the snapshot must be taken before the target's config is applied" + rollback = runtime.split("const rollbackSpeculativeType", 1)[1] + assert "nParallel: previousNParallel," in rollback + # Baseline and reload payload keep the resolved count, or the rollback + # recreates the previous model at a different slot count. + assert "loadedNParallel: stateBeforeUnload.loadedNParallel ?? null," in rollback + assert "n_parallel: stateBeforeUnload.loadedNParallel," in runtime + + def test_vulkan_inference_devices_are_the_pickable_set(): """GGUF loads run through llama-server, so on a Vulkan build the picker must offer the inference inventory (ggml ordinals, the space `--device Vulkan` diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py index 864941a20a..9fd264ddf5 100644 --- a/unsloth_cli/commands/studio.py +++ b/unsloth_cli/commands/studio.py @@ -1263,7 +1263,8 @@ def studio_default( max = _PARALLEL_MAX, help = ( f"llama-server parallel decode slots ({_PARALLEL_MIN}..{_PARALLEL_MAX}). " - f"Default {_PARALLEL_DEFAULT_PLAIN}." + f"Default {_PARALLEL_DEFAULT_PLAIN}. The Studio run settings " + "(Parallel Slots) override it per load." ), ), cloudflare: Optional[bool] = typer.Option( @@ -1880,7 +1881,8 @@ def run( help = ( "llama-server parallel decode slots. N requests share one " "loaded model; each slot gets ctx/N KV cache. Default " - f"{_PARALLEL_DEFAULT_RUN} (pre-PR hardcoded value)." + f"{_PARALLEL_DEFAULT_RUN} (pre-PR hardcoded value). The Studio " + "run settings (Parallel Slots) can override it per load." ), ), cloudflare: Optional[bool] = typer.Option( From ddb93448089d61b911543849bce31a578dd62fa4 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:07:20 -0700 Subject: [PATCH 08/16] Route the stale-manifest abort through Exit-SetupFailure (#7570) From 85c63e790346491d9e14f41fcb4fdf51923164ac Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:08:47 -0700 Subject: [PATCH 09/16] Studio: honour LLAMA_ARG_FLASH_ATTN when recording the launched flash-attention state (#7557) --- studio/backend/core/inference/llama_cpp.py | 18 ++++++++++++++--- studio/backend/tests/test_mtp_vram_budget.py | 21 +++++++++++++++++++- 2 files changed, 35 insertions(+), 4 deletions(-) diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py index 4f2d8cd54a..5b32103dc8 100644 --- a/studio/backend/core/inference/llama_cpp.py +++ b/studio/backend/core/inference/llama_cpp.py @@ -1697,9 +1697,20 @@ def _kv_unified_from_args( return enabled -def _flash_attn_enabled_from_args(args: Optional[Iterable[str]], default: bool = True) -> bool: - """Resolve llama.cpp's last-wins flash-attention CLI setting.""" +def _flash_attn_enabled_from_args( + args: Optional[Iterable[str]], + default: bool = True, + env: Optional[Mapping[str, str]] = None, +) -> bool: + """Resolve llama.cpp's environment and last-wins flash-attention settings.""" enabled = default + # llama.cpp applies LLAMA_ARG_FLASH_ATTN before parsing argv (arg.cpp set_env), + # so the CLI still wins. --flash-attn has no args_neg, so no LLAMA_ARG_NO_ twin. + value = (os.environ if env is None else env).get("LLAMA_ARG_FLASH_ATTN") + if value in _LLAMA_ARG_FALSE_VALUES: + enabled = False + elif value in _LLAMA_ARG_TRUE_OR_AUTO_VALUES: + enabled = True values = [str(arg) for arg in args] if args else [] for i, raw in enumerate(values): if _flag_name(raw) not in {"-fa", "--flash-attn"}: @@ -9054,7 +9065,8 @@ class LlamaCppBackend: int(self._DEFAULT_N_UBATCH if _effective_ubatch is None else _effective_ubatch), ) self._flash_attn_enabled = ( - _flash_attn_enabled_from_args(_last_spawn_cmd) and self._architecture != "grok" + _flash_attn_enabled_from_args(_last_spawn_cmd, env = env) + and self._architecture != "grok" ) self._effective_cache_types = _effective_main_cache_types( _last_spawn_cmd, diff --git a/studio/backend/tests/test_mtp_vram_budget.py b/studio/backend/tests/test_mtp_vram_budget.py index 77ca76325f..3742018e5e 100644 --- a/studio/backend/tests/test_mtp_vram_budget.py +++ b/studio/backend/tests/test_mtp_vram_budget.py @@ -817,7 +817,26 @@ class TestExtraArgsMtpDetection: ], ) def test_flash_attn_last_value_wins(self, args, expected): - assert _flash_attn_enabled_from_args(args) is expected + assert _flash_attn_enabled_from_args(args, env = {}) is expected + + @pytest.mark.parametrize( + "value,expected", + [ + ("off", False), + ("disabled", False), + ("false", False), + ("0", False), + ("on", True), + ("auto", True), + ("garbage", True), # llama.cpp refuses to start, so the default is moot + ], + ) + def test_flash_attn_env_applies(self, value, expected): + env = {"LLAMA_ARG_FLASH_ATTN": value} + assert _flash_attn_enabled_from_args([], env = env) is expected + # llama.cpp parses the environment first, so an explicit flag still wins. + assert _flash_attn_enabled_from_args(["-fa", "on"], env = env) is True + assert _flash_attn_enabled_from_args(["-fa", "off"], env = env) is False def test_effective_main_cache_types_follow_env_then_cli(self): env = { From 411cb86d6223f35d257747e7221b5d06c005f9b1 Mon Sep 17 00:00:00 2001 From: Leo Borcherding Date: Tue, 28 Jul 2026 20:12:26 -0500 Subject: [PATCH 10/16] amd: require bitsandbytes>=0.50.0 in the amd extra (fixes ROCm 4-bit NaNs) (#7535) * amd: require bitsandbytes>=0.50.0 in the amd extra bnb <= 0.49.2 NaNs at decode shape on every AMD GPU. The ROCm 4-bit GEMV fix (bnb PR #1887) first ships in 0.50.0, on PyPI since 2026-07-24, so the old >=0.49.1 floor could still resolve the broken range. Mirrors the same change made on the pip release branch in #7278. * amd: cite the 0.50.0 ROCm work accurately in the bnb floor comment The comment credited bnb PR #1887 as "the ROCm 4-bit GEMV fix" for every AMD GPU. #1887 decouples blocksize from warp size and fixes a hardcoded warp size of 32 in kgemm_4bit_inference_naive, which is a CDNA problem by construction. The RDNA-side work is #1979 (fused 4-bit SIMT GEMM) and #2012 (RDNA3/4 workgroup resonance). All three first ship in 0.50.0, so the >=0.50.0 floor is unchanged; only the justification was wrong. * amd: raise the installer bitsandbytes fallback floors to 0.50.0 * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * amd: stop reporting the bitsandbytes PyPI fallback as broken * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Tighten AMD bnb floor comments * Keep the amd extra citation and the AMD install guide reference * amd: do not promise aarch64 a ROCm 4-bit backend it never gets * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * amd: fall back to the PyPI bitsandbytes floor on Windows ROCm too * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: Daniel Han Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --- install.sh | 40 +++++++-- pyproject.toml | 7 +- studio/install_python_stack.py | 99 +++++++++++++++------- tests/python/test_cross_platform_parity.py | 73 ++++++++++++++++ tests/studio/install/test_rocm_support.py | 31 ++++++- 5 files changed, 205 insertions(+), 45 deletions(-) diff --git a/install.sh b/install.sh index 376daa8fab..72f2455277 100755 --- a/install.sh +++ b/install.sh @@ -321,10 +321,25 @@ _gfx906_bnb_prune() { || "$_VENV_PY" -m pip uninstall -y bitsandbytes >/dev/null 2>&1 || true } -# Install bitsandbytes on AMD ROCm hosts. Uses the continuous-release_main -# wheel for the ROCm 4-bit GEMV fix (bnb PR #1887, post-0.49.2); bnb <= 0.49.2 -# NaNs at decode shape on every AMD GPU. Falls back to PyPI >=0.49.1 if the -# pre-release URL is unreachable. Drop the pin once bnb 0.50+ ships on PyPI. +# Install bitsandbytes on AMD ROCm hosts. bnb <= 0.49.2 NaNs at 4-bit decode +# shape on every AMD GPU; the fix (bnb #1887) ships in continuous-release_main +# and, on PyPI, first in 0.50.0. Keep this floor in step with the amd extra in +# pyproject.toml and studio/install_python_stack.py. +_BNB_ROCM_PYPI_FALLBACK="bitsandbytes>=0.50.0" +# bitsandbytes ships no ROCm binary in its aarch64 wheel at any version: the PyPI +# 0.50.0 and continuous-release_main aarch64 wheels both carry only +# libbitsandbytes_cpu.so plus CUDA variants. So neither install path below gives +# aarch64 a 4-bit backend, and the messages must not claim one. Cf. gfx906. +_bnb_rocm_arch_has_binary() { + case "$_ARCH" in + aarch64|arm64) return 1 ;; + *) return 0 ;; + esac +} +_warn_bnb_no_rocm_binary() { + _bnb_rocm_arch_has_binary && return 0 + substep "[WARN] aarch64: bitsandbytes ships no ROCm kernels on this arch; 4-bit QLoRA needs a source build -- https://docs.unsloth.ai/get-started/install-and-update/amd" "$C_WARN" +} _install_bnb_rocm() { _label="$1" _venv_py="$2" @@ -339,9 +354,8 @@ _install_bnb_rocm() { _bnb_whl_url="" ;; esac - # uv rejects the continuous-release_main bitsandbytes wheel because the - # filename version (1.33.7rc0) does not match the embedded metadata version - # (0.50.0.dev0). pip accepts the mismatch, so bootstrap pip and use it. + # uv rejects the pre-release wheel: filename version (1.33.7rc0) does not + # match metadata (0.50.x.dev0). pip accepts it, so bootstrap pip and use it. if ! "$_venv_py" -m pip --version >/dev/null 2>&1; then if ! run_maybe_quiet "$_venv_py" -m ensurepip --upgrade; then run_maybe_quiet uv pip install --python "$_venv_py" pip || \ @@ -357,6 +371,7 @@ _install_bnb_rocm() { --retries 8 --timeout 90 \ "$_bnb_whl_url" >"$_bnb_log" 2>&1; then rm -f "$_bnb_log" + _warn_bnb_no_rocm_binary return 0 fi _bnb_rc=$? @@ -365,10 +380,17 @@ _install_bnb_rocm() { fi rm -f "$_bnb_log" step "warning" "$_label (pre-release) failed (exit code $_bnb_rc)" "$C_WARN" >&2 - substep "[WARN] bnb pre-release install failed; falling back to PyPI (4-bit decode broken on ROCm)" "$C_WARN" + if _bnb_rocm_arch_has_binary; then + substep "[WARN] bnb pre-release install failed; falling back to PyPI $_BNB_ROCM_PYPI_FALLBACK, which carries the ROCm 4-bit fix" "$C_WARN" + else + substep "[WARN] bnb pre-release install failed; falling back to PyPI $_BNB_ROCM_PYPI_FALLBACK" "$C_WARN" + fi fi run_install_cmd "$_label (pypi fallback)" "$_venv_py" -m pip install \ - --force-reinstall --no-cache-dir --no-deps "bitsandbytes>=0.49.1" + --force-reinstall --no-cache-dir --no-deps "$_BNB_ROCM_PYPI_FALLBACK" + _bnb_pypi_rc=$? + _warn_bnb_no_rocm_binary + return $_bnb_pypi_rc } if [ "$_next_is_package" = true ]; then diff --git a/pyproject.toml b/pyproject.toml index 62623499d6..7359a51fa6 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1257,8 +1257,11 @@ intel = [ ] amd = [ "unsloth[huggingfacenotorch]", - "bitsandbytes>=0.49.1 ; ('linux' in sys_platform) and (platform_machine == 'AMD64' or platform_machine == 'x86_64' or platform_machine == 'aarch64')", - "bitsandbytes>=0.49.1 ; (sys_platform == 'win32') and (platform_machine == 'AMD64' or platform_machine == 'x86_64')", + # 4-bit decode is unreliable on ROCm before 0.50.0, the first PyPI release + # carrying the full path: blocksize/warp decoupling (bnb #1887), fused SIMT + # GEMM on RDNA (#1979), RDNA3/4 workgroup fix (#2012). + "bitsandbytes>=0.50.0 ; ('linux' in sys_platform) and (platform_machine == 'AMD64' or platform_machine == 'x86_64' or platform_machine == 'aarch64')", + "bitsandbytes>=0.50.0 ; (sys_platform == 'win32') and (platform_machine == 'AMD64' or platform_machine == 'x86_64')", ] rocm702-torch280 = [ "unsloth[amd]", diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py index 8c71d39e16..3243089656 100644 --- a/studio/install_python_stack.py +++ b/studio/install_python_stack.py @@ -426,8 +426,8 @@ _GFX_TO_AMD_INDEX_ARCH: dict[str, str] = { } # bitsandbytes continuous-release_main wheels with the ROCm 4-bit GEMV fix -# (bnb PR #1887, post-0.49.2). bnb <= 0.49.2 NaNs at decode shape on every -# AMD GPU. Drop the pin once bnb 0.50+ ships on PyPI. +# (bnb #1887, post-0.49.2). bnb <= 0.49.2 NaNs at decode shape on every AMD GPU; +# PyPI 0.50.0 is the first release with the fix, so the fallback below is safe. _BNB_ROCM_PRERELEASE_URLS: dict[str, str] = { "x86_64": ( "https://github.com/bitsandbytes-foundation/bitsandbytes/releases/" @@ -448,7 +448,8 @@ _BNB_ROCM_PRERELEASE_URLS: dict[str, str] = { "bitsandbytes-1.33.7.preview-py3-none-win_amd64.whl" ), } -_BNB_ROCM_PYPI_FALLBACK = "bitsandbytes>=0.49.1" +# Keep in step with the amd extra in pyproject.toml and the install.sh fallback. +_BNB_ROCM_PYPI_FALLBACK = "bitsandbytes>=0.50.0" def _bnb_rocm_prerelease_url() -> str | None: @@ -460,6 +461,16 @@ def _bnb_rocm_prerelease_url() -> str | None: return _BNB_ROCM_PRERELEASE_URLS.get(arch) +def _bnb_rocm_arch_has_binary() -> bool: + """False on aarch64: bitsandbytes ships no ROCm kernels there at any version. + The PyPI 0.50.0 and continuous-release_main aarch64 wheels both carry only + libbitsandbytes_cpu.so plus CUDA variants, so neither install path gives + aarch64 a 4-bit backend and neither message may claim one. + """ + arch = platform.machine().lower() + return {"amd64": "x86_64", "arm64": "aarch64"}.get(arch, arch) != "aarch64" + + def _amd_smi_env() -> dict[str, str] | None: """On Windows, env with __COMPAT_LAYER=RunAsInvoker; None elsewhere. NB: RunAsInvoker doesn't stop amd-smi's runtime elevation (its manifest is @@ -1243,29 +1254,46 @@ _rocm_windows_torch_installed: bool = False def _install_bnb_windows_rocm() -> bool: - """Install the AMD Windows BNB prerelease wheel. Returns True on success. + """Install AMD Windows BNB, pre-release wheel first. Returns True on success. - The continuous-release wheel is intentionally mismatched: the filename - encodes 1.33.7.preview (parsed as 1.33.7rc0 by PEP 440) while the wheel - metadata reports 0.50.0.dev0. uv rejects this filename/metadata mismatch, - and bypassing it with UV_SKIP_WHEEL_FILENAME_CHECK still leaves uv mangling - the bitsandbytes install. Per the AMD install guide - (https://unsloth.ai/docs/get-started/install/amd/amd-hackathon) the wheel - must be installed with plain pip, not uv, so we force pip (force_pip=True); - plain pip performs no wheel filename/metadata check. + The wheel's filename version (1.33.7.preview, PEP 440 1.33.7rc0) does not + match its metadata (0.50.x.dev0). uv rejects the mismatch and still mangles + the install under UV_SKIP_WHEEL_FILENAME_CHECK, so force plain pip, which + performs no such check. Per the AMD install guide + (https://unsloth.ai/docs/get-started/install/amd/amd-hackathon). + + When that URL is blocked, fall back to PyPI. Its win_amd64 wheel ships + libbitsandbytes_rocm{714,72}.dll from 0.50.0 on, so the fallback is a real + ROCm build; before 0.50.0 it was CUDA-only, which is why there was none. """ _bnb_win_url = _BNB_ROCM_PRERELEASE_URLS.get("win_amd64") - if _bnb_win_url is None: - return False - _ok = pip_install_try( - "bitsandbytes (AMD Windows, pre-release main)", - "--force-reinstall", - "--no-cache-dir", - "--no-deps", - _bnb_win_url, - constrain = False, - force_pip = True, - ) + _ok = False + if _bnb_win_url is not None: + _ok = pip_install_try( + "bitsandbytes (AMD Windows, pre-release main)", + "--force-reinstall", + "--no-cache-dir", + "--no-deps", + _bnb_win_url, + constrain = False, + force_pip = True, + ) + if not _ok: + print( + _red( + " bnb pre-release install failed; falling back to PyPI " + f"{_BNB_ROCM_PYPI_FALLBACK}, which carries the ROCm 4-bit fix" + ) + ) + if not _ok: + _ok = pip_install_try( + "bitsandbytes (AMD Windows)", + "--force-reinstall", + "--no-cache-dir", + "--no-deps", + _BNB_ROCM_PYPI_FALLBACK, + constrain = False, + ) if not _ok: return False # Detect the actual ROCm DLL suffix in the wheel and set BNB_ROCM_VERSION so bnb @@ -1755,8 +1783,8 @@ def _ensure_rocm_torch() -> None: pass if _torch_ok: _rocm_windows_torch_installed = True - # ROCm torch is already installed, but the AMD Windows BNB wheel is still - # needed (the PyPI bitsandbytes ships only CUDA DLLs, fails on ROCm). + # ROCm torch is already installed, but bnb still needs the ROCm build + # (pre-release wheel, else PyPI >=0.50.0). _install_bnb_windows_rocm() return # torch was wiped between runs; fall through to the full install path @@ -1834,12 +1862,12 @@ def _ensure_rocm_torch() -> None: # separate dependency -- a BNB install failure must NOT roll back the # torch ROCm install. _rocm_windows_torch_installed = True - # Always install AMD Windows bitsandbytes -- the PyPI wheel ships only - # CUDA DLLs and fails on ROCm. Install even when torch was already a - # ROCm build so `studio update` repairs a broken bnb. + # Always install AMD Windows bitsandbytes, even when torch was already a + # ROCm build, so `studio update` repairs a broken bnb. if not _install_bnb_windows_rocm(): print( - " Warning: AMD Windows bitsandbytes install failed; " + " Warning: AMD Windows bitsandbytes install failed " + "(pre-release and PyPI); " "ROCm torch is installed but bitsandbytes may need manual install" ) return @@ -2170,10 +2198,13 @@ def _ensure_rocm_torch() -> None: force_pip = True, ) if not _bnb_installed: + _fallback_note = ( + ", which carries the ROCm 4-bit fix" if _bnb_rocm_arch_has_binary() else "" + ) print( _red( " bnb pre-release install failed; falling back to PyPI " - "(4-bit decode will be broken on ROCm)" + f"{_BNB_ROCM_PYPI_FALLBACK}{_fallback_note}" ) ) if not _bnb_installed: @@ -2185,6 +2216,14 @@ def _ensure_rocm_torch() -> None: _BNB_ROCM_PYPI_FALLBACK, constrain = False, ) + if not _bnb_rocm_arch_has_binary(): + print( + _red( + " aarch64: bitsandbytes ships no ROCm kernels on this arch; " + "4-bit QLoRA needs a source build -- " + "https://docs.unsloth.ai/get-started/install-and-update/amd" + ) + ) # _uv_safe_path is imported from backend.utils.uv_path_safety (shared with mlx_repair). diff --git a/tests/python/test_cross_platform_parity.py b/tests/python/test_cross_platform_parity.py index 6c2a1d09cf..b20e715ebc 100644 --- a/tests/python/test_cross_platform_parity.py +++ b/tests/python/test_cross_platform_parity.py @@ -862,3 +862,76 @@ class TestNoTorchPersistenceParity: manifest = (REPO_ROOT / "studio" / "install_manifest.py").read_text(encoding = "utf-8") assert 'NO_TORCH_TRUTHY: Tuple[str, ...] = ("1", "true", "yes", "on")' in manifest assert "install_manifest.NO_TORCH_TRUTHY" in STACK_PY.read_text(encoding = "utf-8") + + +class TestAmdBnbFloorParity: + """bitsandbytes <= 0.49.2 NaNs at 4-bit decode shape on every AMD GPU; the ROCm + 4-bit GEMV fix (bnb #1887) first ships on PyPI in 0.50.0. The `amd` extra, + install.sh and the Studio stack resolve bitsandbytes independently, so all three + must carry the same floor or an unreachable pre-release wheel silently reinstates + the broken range.""" + + FLOOR = "0.50.0" + PYPROJECT = REPO_ROOT / "pyproject.toml" + + def test_amd_extra_floor(self): + text = self.PYPROJECT.read_text(encoding = "utf-8") + amd = re.search(r"^amd = \[(.*?)^\]", text, re.S | re.M) + assert amd, "pyproject.toml must define an `amd` extra" + specs = re.findall(r'"(bitsandbytes[^"]*)"', amd.group(1)) + assert specs, "the amd extra must pin bitsandbytes" + for spec in specs: + assert spec.startswith( + f"bitsandbytes>={self.FLOOR}" + ), f"amd extra bitsandbytes floor must be >={self.FLOOR}, got {spec!r}" + + def test_install_sh_pypi_fallback_floor(self): + text = INSTALL_SH.read_text(encoding = "utf-8") + assert ( + f'_BNB_ROCM_PYPI_FALLBACK="bitsandbytes>={self.FLOOR}"' in text + ), f"install.sh _install_bnb_rocm PyPI fallback must floor at {self.FLOOR}" + + def test_stack_py_pypi_fallback_floor(self): + text = STACK_PY.read_text(encoding = "utf-8") + assert ( + f'_BNB_ROCM_PYPI_FALLBACK = "bitsandbytes>={self.FLOOR}"' in text + ), f"install_python_stack.py PyPI fallback must floor at {self.FLOOR}" + + def test_no_installer_still_allows_the_broken_range(self): + for path in (INSTALL_SH, INSTALL_PS1, SETUP_PS1, STACK_PY, self.PYPROJECT): + text = path.read_text(encoding = "utf-8") + for line in text.splitlines(): + if "bitsandbytes>=0.49" in line and not line.lstrip().startswith(("#", "//")): + raise AssertionError( + f"{path.name} still floors bitsandbytes in the broken ROCm range: {line.strip()!r}" + ) + + def test_fallback_is_not_reported_as_broken(self): + """The fallback now installs the first fixed release, so neither installer + may still call 4-bit decode broken on ROCm.""" + for path in (INSTALL_SH, STACK_PY): + text = path.read_text(encoding = "utf-8") + assert ( + "4-bit decode broken on ROCm" not in text + ), f"{path.name} still reports the repaired PyPI fallback as broken" + assert ( + "4-bit decode will be broken on ROCm" not in text + ), f"{path.name} still reports the repaired PyPI fallback as broken" + + def test_aarch64_is_not_told_it_has_a_rocm_backend(self): + """bitsandbytes ships no ROCm kernels in its aarch64 wheel at any version, so + neither installer may hand aarch64 the x86_64 "carries the ROCm 4-bit fix" + message, and both must warn that 4-bit needs a source build there.""" + sh = INSTALL_SH.read_text(encoding = "utf-8") + assert "_bnb_rocm_arch_has_binary()" in sh + assert "_warn_bnb_no_rocm_binary()" in sh + assert ( + sh.count("_warn_bnb_no_rocm_binary\n") >= 2 + ), "install.sh must warn on aarch64 after both the pre-release and the fallback install" + py = STACK_PY.read_text(encoding = "utf-8") + assert "def _bnb_rocm_arch_has_binary(" in py + assert "_bnb_rocm_arch_has_binary()" in py + for text, name in ((sh, "install.sh"), (py, "install_python_stack.py")): + assert ( + "4-bit QLoRA needs a source build" in text + ), f"{name} must tell aarch64 users 4-bit needs a source build" diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py index b003382859..51c2d6587c 100644 --- a/tests/studio/install/test_rocm_support.py +++ b/tests/studio/install/test_rocm_support.py @@ -3157,12 +3157,35 @@ class TestInstallBnbWindowsRocm: assert result is False assert "BNB_ROCM_VERSION" not in os.environ - def test_no_op_when_win_amd64_url_missing(self): - """Should be silent no-op if win_amd64 key absent from _BNB_ROCM_PRERELEASE_URLS.""" + def test_falls_back_to_pypi_when_win_amd64_url_missing(self): + """No win_amd64 pre-release wheel must not mean no bitsandbytes: PyPI + >=0.50.0 ships libbitsandbytes_rocm{714,72}.dll, so it is a real ROCm build.""" with patch.object(stack_mod, "_BNB_ROCM_PRERELEASE_URLS", {}): - with patch.object(stack_mod, "pip_install_try") as mock_pip: + with patch.object(stack_mod, "pip_install_try", return_value = True) as mock_pip: stack_mod._install_bnb_windows_rocm() - mock_pip.assert_not_called() + assert mock_pip.call_count == 1 + assert stack_mod._BNB_ROCM_PYPI_FALLBACK in mock_pip.call_args.args + + def test_falls_back_to_pypi_when_prerelease_install_fails(self): + """A blocked GitHub pre-release URL must fall through to the PyPI floor rather + than leaving Windows ROCm with no working bitsandbytes.""" + with patch.object(stack_mod, "pip_install_try", side_effect = [False, True]) as mock_pip: + with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver", return_value = "72"): + result = stack_mod._install_bnb_windows_rocm() + assert result is True + assert mock_pip.call_count == 2 + assert "win_amd64" in str(mock_pip.call_args_list[0]) + assert stack_mod._BNB_ROCM_PYPI_FALLBACK in mock_pip.call_args_list[1].args + + def test_returns_false_only_when_both_paths_fail(self): + """Both the pre-release wheel and the PyPI fallback must fail before the + helper reports failure.""" + with patch.dict(os.environ, {}, clear = False): + os.environ.pop("BNB_ROCM_VERSION", None) + with patch.object(stack_mod, "pip_install_try", return_value = False) as mock_pip: + result = stack_mod._install_bnb_windows_rocm() + assert result is False + assert mock_pip.call_count == 2 def test_sets_bnb_rocm_version_from_detected_dll(self): """BNB_ROCM_VERSION is set from the DLL detected after install.""" From a0a3a7b24a3bcf4383e66f89782f547e8f5071bb Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:19:39 -0700 Subject: [PATCH 11/16] fix(studio): show the current artifact's source after switching artifacts (#7565) * fix(studio): show the current artifact's source after switching artifacts The canvas source view feeds one Streamdown a fence built from the selected artifact's code, but never keys it. Streamdown does not revise a block it has already committed, so the panel keeps rendering the previous artifact's source. Key the source view on the artifact ID plus a hash of its code: tool artifact IDs are derived from the tool call, not the code, so the ID alone does not change when a tool artifact is updated in place. * Name the real root cause and make the source-key test load-bearing The remount is needed because Streamdown memoizes a fenced code block on its hast node's line/column span, which ignores the text inside the fence, so two canvases of equal line count compare equal and the old source stays on screen. Verified in Chromium against streamdown 2.5.0: unkeyed, 70 lines -> 70 lines renders the previous artifact, 70 -> 71 and 70 -> 90 render correctly. Move the key expression into the source branch so it costs nothing while the artifact is streaming and the view is unmounted, and export the helper from types.ts so the test exercises the shipped code instead of a local copy of the formula (it passed before even with the key removed from the component). * Assert the source view's Streamdown key wiring, not just the helper The suite exercised buildArtifactSourceKey but never the component, so deleting key={buildArtifactSourceKey(artifact)} from the Streamdown left every test green. There is no DOM renderer available to these tests, so parse artifact-surface.tsx with the TypeScript compiler API (already a devDependency) and assert the source view's Streamdown carries that key. Mutation-checked: removing the key fails 1 test, swapping it for artifact.id fails 1, and making the helper ignore code fails 2. * Tighten the comments added by this PR --- .../chat/artifacts/artifact-surface.tsx | 4 +- .../src/features/chat/artifacts/types.ts | 9 ++ .../tests/artifact-source-key.test.ts | 130 ++++++++++++++++++ 3 files changed, 142 insertions(+), 1 deletion(-) create mode 100644 studio/frontend/tests/artifact-source-key.test.ts diff --git a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx index 1955c3aca1..4e28e7f457 100644 --- a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx +++ b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx @@ -30,7 +30,7 @@ import { Streamdown } from "streamdown"; import { ArtifactHtmlFrame, type ArtifactViewMode } from "./html-frame"; import { useChatArtifactsStore } from "./store"; import type { ChatArtifact } from "./types"; -import { getArtifactFilename } from "./types"; +import { buildArtifactSourceKey, getArtifactFilename } from "./types"; const COPY_RESET_MS = 2000; const artifactSourceCodePlugin = createCodePlugin({ @@ -338,6 +338,8 @@ export function ArtifactSurface({ ) : (
>> 0).toString(36); } +// The canvas source view keys its Streamdown on this. Streamdown memoizes a code +// fence on its node's line/column span, ignoring the text, so equal-line-count +// canvases keep the old source. Tool artifact IDs omit the code, so hash it in. +export function buildArtifactSourceKey( + artifact: Pick, +): string { + return `${artifact.id}:${hashArtifactCode(artifact.code)}`; +} + export function createArtifactId(input: ChatArtifactInput): string { const threadSegment = input.threadId || "no-thread"; const messageSegment = input.sourceMessageId || "transient"; diff --git a/studio/frontend/tests/artifact-source-key.test.ts b/studio/frontend/tests/artifact-source-key.test.ts new file mode 100644 index 0000000000..e90037e603 --- /dev/null +++ b/studio/frontend/tests/artifact-source-key.test.ts @@ -0,0 +1,130 @@ +// SPDX-License-Identifier: AGPL-3.0-only +// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +import assert from "node:assert/strict"; +import { readFileSync } from "node:fs"; +import test from "node:test"; +import { fileURLToPath } from "node:url"; + +import ts from "typescript"; + +import { + buildArtifactSourceKey, + createArtifactId, + createChatArtifact, + hashArtifactCode, +} from "../src/features/chat/artifacts/types.ts"; + +// The shipped helper the component keys on, not a copy of it. +const sourceKey = buildArtifactSourceKey; + +const toolInput = (code: string) => ({ + code, + source: "tool" as const, + threadId: "thread-1", + sourceMessageId: "msg-1", + sourceToolCallId: "call_0", +}); + +const fenceInput = (code: string) => ({ + code, + source: "fence" as const, + threadId: "thread-1", + sourceMessageId: "msg-1", +}); + +test("tool artifact IDs are stable across code changes, so the ID alone is not enough", () => { + const first = createArtifactId(toolInput("

first

")); + const second = createArtifactId(toolInput("

second

")); + assert.equal(first, second); +}); + +test("the source key changes when a tool artifact's code changes", () => { + const first = createChatArtifact(toolInput("

first

")); + const second = createChatArtifact(toolInput("

second

")); + assert.notEqual(sourceKey(first), sourceKey(second)); +}); + +test("the source key changes when switching between fence artifacts", () => { + const first = createChatArtifact(fenceInput("

alpha

")); + const second = createChatArtifact(fenceInput("

bravo

")); + assert.notEqual(sourceKey(first), sourceKey(second)); +}); + +test("the source key is stable for an unchanged artifact, so no needless remount", () => { + const code = "

same

"; + assert.equal( + sourceKey(createChatArtifact(toolInput(code))), + sourceKey(createChatArtifact(toolInput(code))), + ); +}); + +// Equal line count, the shape where Streamdown's comparator sees no change. +test("the source key changes for two canvases with the same shape", () => { + const first = createChatArtifact( + toolInput("\n\n

Alpha

\n\n"), + ); + const second = createChatArtifact( + toolInput("\n\n

Bravo

\n\n"), + ); + assert.equal(first.code.length, second.code.length); + assert.equal(first.code.split("\n").length, second.code.split("\n").length); + assert.notEqual(sourceKey(first), sourceKey(second)); +}); + +test("hashArtifactCode separates same-length codes and empty from whitespace", () => { + assert.notEqual(hashArtifactCode("

ab

"), hashArtifactCode("

ba

")); + assert.notEqual(hashArtifactCode(""), hashArtifactCode(" ")); +}); + +const KEYED_BY_HELPER = /^\{buildArtifactSourceKey\(\s*artifact\s*\)\}$/; + +const SURFACE_PATH = fileURLToPath( + new URL( + "../src/features/chat/artifacts/artifact-surface.tsx", + import.meta.url, + ), +); + +/** The opening tag of `node`, for both `` and ``. */ +const openingTag = (node: ts.Node): ts.JsxOpeningLikeElement | null => { + if (ts.isJsxSelfClosingElement(node)) return node; + if (ts.isJsxElement(node)) return node.openingElement; + return null; +}; + +/** The `key` expression on the source view's Streamdown, or null if unkeyed. */ +function readStreamdownKey(): string | null { + const source = ts.createSourceFile( + SURFACE_PATH, + readFileSync(SURFACE_PATH, "utf8"), + ts.ScriptTarget.ESNext, + true, + ts.ScriptKind.TSX, + ); + let key: string | null = null; + const visit = (node: ts.Node): void => { + const opening = openingTag(node); + if (opening?.tagName.getText() === "Streamdown") { + for (const attribute of opening.attributes.properties) { + if ( + ts.isJsxAttribute(attribute) && + attribute.name.getText() === "key" + ) { + key = attribute.initializer?.getText() ?? ""; + } + } + } + node.forEachChild(visit); + }; + source.forEachChild(visit); + return key; +} + +// Without this the suite passes with the key deleted, which is the regression. +// No DOM renderer is available here, so assert the wiring in the source. +test("the source view's Streamdown is keyed by the shipped helper", () => { + const key = readStreamdownKey(); + assert.ok(key, "source view has no key prop"); + assert.match(key, KEYED_BY_HELPER); +}); From 570c80478541b594ddfd65041eaa297cf1346364 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:20:50 -0700 Subject: [PATCH 12/16] Studio: surface the tool-call nudge in the chat UI (#7559) * Studio: show a Nudging tool calls badge while the tool-call re-prompt runs * Guard the nudge status ordering assertion against index 0 * Tighten the nudge status comments * Announce the nudge text instead of the generic spinner label * Trim the nudge status comments Collapse the multi-line notes to fewer lines and drop one that restated the assert below it. The blank-before-badge ordering reason and the keep-in-sync contract are preserved. --------- Co-authored-by: danielhanchen --- studio/backend/core/inference/llama_cpp.py | 4 + .../core/inference/safetensors_agentic.py | 6 +- .../core/inference/tool_call_parser.py | 3 + .../backend/tests/test_llama_cpp_tool_loop.py | 135 ++++++++++++++++++ .../tests/test_safetensors_tool_loop.py | 19 +++ .../src/components/assistant-ui/thread.tsx | 22 ++- .../src/features/chat/utils/tool-status.ts | 15 ++ studio/frontend/tests/tool-status.test.ts | 45 ++++++ 8 files changed, 243 insertions(+), 6 deletions(-) create mode 100644 studio/frontend/src/features/chat/utils/tool-status.ts create mode 100644 studio/frontend/tests/tool-status.test.ts diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py index 5b32103dc8..144aa1fd37 100644 --- a/studio/backend/core/inference/llama_cpp.py +++ b/studio/backend/core/inference/llama_cpp.py @@ -92,6 +92,7 @@ from utils.subprocess_compat import ( from utils.process_lifetime import child_popen_kwargs as _child_popen_kwargs from core.inference.tool_call_parser import ( MAX_ACT_REPROMPTS as _MAX_REPROMPTS, + NUDGE_TOOL_CALLS_STATUS as _NUDGE_TOOL_CALLS_STATUS, REPROMPT_MAX_CHARS as _REPROMPT_MAX_CHARS, is_short_intent_without_action as _is_short_intent_without_action, reprompt_to_act_message as _reprompt_to_act_message, @@ -12419,7 +12420,10 @@ class LlamaCppBackend: _it_r = _iter_timings or {} _accumulated_predicted_ms += _it_r.get("predicted_ms", 0) _accumulated_predicted_n += _it_r.get("predicted_n", 0) + # Blank first (the route resets its text cursor only on an + # empty status), then the badge so the retry is not a hang. yield {"type": "status", "text": ""} + yield {"type": "status", "text": _NUDGE_TOOL_CALLS_STATUS} continue if _forced_tool_call_pending: diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py index b593bc119b..3057f7c2ac 100644 --- a/studio/backend/core/inference/safetensors_agentic.py +++ b/studio/backend/core/inference/safetensors_agentic.py @@ -35,6 +35,7 @@ from core.inference.tool_call_parser import ( _strip_mistral_reasoning, BUDGET_EXHAUSTED_NUDGE, MAX_ACT_REPROMPTS, + NUDGE_TOOL_CALLS_STATUS, RAG_MAX_SEARCHES_PER_TURN, RAG_SEARCH_CAP_NUDGE, TOOL_XML_SIGNALS, @@ -1032,9 +1033,10 @@ def run_safetensors_tool_loop( "content": reprompt_to_act_message(tool_hint), } ) - # Empty status clears the badge and resets the route's - # per-turn text cursor before the re-prompted turn streams. + # Blank first: it clears the badge and resets the route's per-turn + # text cursor. The badge then shows the pause is a re-prompt, not a stall. yield {"type": "status", "text": ""} + yield {"type": "status", "text": NUDGE_TOOL_CALLS_STATUS} continue # Final answer. If a literal tool marker in prose was buffered but diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py index 9b6b0a7773..4c3fe234ae 100644 --- a/studio/backend/core/inference/tool_call_parser.py +++ b/studio/backend/core/inference/tool_call_parser.py @@ -183,6 +183,9 @@ INTENT_SIGNAL = re.compile( # times since #5620); safetensors and MLX inherit the same cap from here. MAX_ACT_REPROMPTS = 3 REPROMPT_MAX_CHARS = 2000 +# Composer badge while a hidden re-prompted turn regenerates, else the UI looks +# hung. Matched exactly by the frontend (utils/tool-status.ts); keep in sync. +NUDGE_TOOL_CALLS_STATUS = "Nudging tool calls" def is_short_intent_without_action(text: str) -> bool: diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py index c629ff3be4..cbd1b07505 100644 --- a/studio/backend/tests/test_llama_cpp_tool_loop.py +++ b/studio/backend/tests/test_llama_cpp_tool_loop.py @@ -26,6 +26,7 @@ from core.inference.llama_cpp import ( _PROVISIONAL_ARGS_MIN_CHARS, LlamaCppBackend, ) +from core.inference.tool_call_parser import NUDGE_TOOL_CALLS_STATUS from state import tool_approvals from state.tool_approvals import TOOL_REJECTED_MESSAGE, resolve_tool_decision @@ -1841,6 +1842,140 @@ def test_reprompted_tool_call_still_streams_final_answer(monkeypatch): assert len(payloads) == 3 +def _status_texts(events: list[dict]) -> list[str]: + return [event["text"] for event in events if event.get("type") == "status"] + + +_WEB_SEARCH_TOOL = { + "type": "function", + "function": { + "name": "web_search", + "description": "Search the web.", + "parameters": { + "type": "object", + "properties": {"query": {"type": "string"}}, + "required": ["query"], + }, + }, +} + + +def _nudge_then_search_streams() -> list[list[str]]: + """Stall, then a re-prompted turn that finally searches, then the answer.""" + + return [ + [_sse({"content": "I will search the web now."}), _done()], + [ + _sse( + { + "tool_calls": [ + { + "index": 0, + "id": "call_search", + "type": "function", + "function": { + "name": "web_search", + "arguments": json.dumps({"query": "red square"}), + }, + } + ] + } + ), + _done(), + ], + [_sse({"content": "Final answer: the square is red."}), _done()], + ] + + +def test_plan_without_action_nudge_is_announced_on_the_status_channel(monkeypatch): + """The re-prompted turn is hidden, so without a badge the UI looks frozen.""" + + payloads: list[dict] = [] + backend = _make_backend(monkeypatch, _nudge_then_search_streams(), payloads) + monkeypatch.setattr( + "core.inference.tools.execute_tool", + lambda *_a, **_k: "Search results: red is #f00.", + ) + + events = list( + backend.generate_chat_completion_with_tools( + messages = [{"role": "user", "content": "What colour is the square?"}], + tools = [_WEB_SEARCH_TOOL], + max_tool_iterations = 2, + ) + ) + + statuses = _status_texts(events) + assert NUDGE_TOOL_CALLS_STATUS in statuses + index = statuses.index(NUDGE_TOOL_CALLS_STATUS) + # Blank first: the route resets its text cursor only on an empty status. + # index > 0 matters: at 0, statuses[-1] wraps to the terminal clear. + assert index > 0 and statuses[index - 1] == "" + assert statuses[index + 1].startswith("Searching:") + assert statuses[-1] == "" + + +def test_plan_without_action_nudge_status_clears_when_the_retry_just_answers(monkeypatch): + streams = [ + [_sse({"content": "I will search the web now."}), _done()], + [_sse({"content": "No search needed. Final answer: the square is red."}), _done()], + ] + payloads: list[dict] = [] + backend = _make_backend(monkeypatch, streams, payloads) + + events = list( + backend.generate_chat_completion_with_tools( + messages = [{"role": "user", "content": "What colour is the square?"}], + tools = [_WEB_SEARCH_TOOL], + max_tool_iterations = 2, + ) + ) + + statuses = _status_texts(events) + assert NUDGE_TOOL_CALLS_STATUS in statuses + assert statuses[-1] == "" + + +def test_direct_answer_never_shows_the_nudge_status(monkeypatch): + payloads: list[dict] = [] + backend = _make_backend( + monkeypatch, + [[_sse({"content": "The square is red."}), _done()]], + payloads, + ) + + events = list( + backend.generate_chat_completion_with_tools( + messages = [{"role": "user", "content": "What colour is the square?"}], + tools = [_WEB_SEARCH_TOOL], + max_tool_iterations = 2, + ) + ) + + assert NUDGE_TOOL_CALLS_STATUS not in _status_texts(events) + + +def test_nudge_status_absent_when_nudging_is_disabled(monkeypatch): + payloads: list[dict] = [] + backend = _make_backend(monkeypatch, _nudge_then_search_streams(), payloads) + monkeypatch.setattr( + "core.inference.tools.execute_tool", + lambda *_a, **_k: "Search results: red is #f00.", + ) + + events = list( + backend.generate_chat_completion_with_tools( + messages = [{"role": "user", "content": "What colour is the square?"}], + tools = [_WEB_SEARCH_TOOL], + max_tool_iterations = 2, + nudge_tool_calls = False, + ) + ) + + assert NUDGE_TOOL_CALLS_STATUS not in _status_texts(events) + assert len(payloads) == 1 + + def test_confirm_tool_calls_allow_executes_gguf_tool(monkeypatch): streams = [ _structured_tool_call("python", {"code": "print(1)"}, "call_py"), diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py index 1043005f64..2e7e99fbba 100644 --- a/studio/backend/tests/test_safetensors_tool_loop.py +++ b/studio/backend/tests/test_safetensors_tool_loop.py @@ -24,6 +24,7 @@ from core.inference.safetensors_agentic import ( strip_tool_markup_streaming, ) from core.inference.tool_call_parser import ( + NUDGE_TOOL_CALLS_STATUS, RAG_MAX_SEARCHES_PER_TURN, has_tool_signal, parse_tool_calls_from_text, @@ -2231,6 +2232,24 @@ def test_reprompt_names_only_active_tools_not_hardcoded(): assert "python" not in reprompt["content"] +def test_reprompt_is_announced_on_the_status_channel(): + # The re-prompted turn is hidden, so the badge is the only sign of life. + # Blank still comes first: the route resets its text cursor only on that. + _captured, events = _reprompt_loop(auto_heal_tool_calls = True) + statuses = [e["text"] for e in events if e["type"] == "status"] + assert NUDGE_TOOL_CALLS_STATUS in statuses + index = statuses.index(NUDGE_TOOL_CALLS_STATUS) + # index > 0 matters: at 0, statuses[-1] wraps to the terminal clear. + assert index > 0 and statuses[index - 1] == "" + assert statuses[-1] == "" + + +def test_reprompt_status_absent_without_a_nudge(): + _captured, events = _reprompt_loop(auto_heal_tool_calls = False) + statuses = [e["text"] for e in events if e["type"] == "status"] + assert NUDGE_TOOL_CALLS_STATUS not in statuses + + def test_reprompt_suppressed_when_auto_heal_disabled(): # With Auto-Heal off the safetensors nudge must stay silent for backend parity # with the GGUF loop, so only the single initial generation runs. diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx index 6da8126421..9b3c7aa79e 100644 --- a/studio/frontend/src/components/assistant-ui/thread.tsx +++ b/studio/frontend/src/components/assistant-ui/thread.tsx @@ -91,6 +91,7 @@ import { useResearchRunStore, } from "@/features/chat/stores/research-run-store"; import { parseExternalModelId } from "@/features/chat/external-providers"; +import { toolStatusKind } from "@/features/chat/utils/tool-status"; import { McpComposerButton } from "@/features/chat/mcp-composer-button"; import { getExternalReasoningCapabilities } from "@/features/chat/provider-capabilities"; import { useRagToolDisabled } from "@/features/chat/hooks/use-rag-tool-disabled"; @@ -2847,15 +2848,28 @@ const ToolStatusDisplay: FC = () => { } // From the store's start time, so returning to the conversation resumes rather than restarting. const elapsed = Math.max(0, Math.floor((now - startedAt) / 1000)); - const isRunning = toolStatus.startsWith("Running"); - const StatusIcon = isRunning ? TerminalIcon : GlobeIcon; + const kind = toolStatusKind(toolStatus); + const isNudging = kind === "nudge"; + const StatusIcon = kind === "terminal" ? TerminalIcon : GlobeIcon; return (
-
- +
+ {isNudging ? ( + // label, not the default "Loading": the spinner is the badge's only + // role="status" region, so its name is what gets announced. + + ) : ( + + )} {toolStatus} {elapsed}s
diff --git a/studio/frontend/src/features/chat/utils/tool-status.ts b/studio/frontend/src/features/chat/utils/tool-status.ts new file mode 100644 index 0000000000..16c86bbd49 --- /dev/null +++ b/studio/frontend/src/features/chat/utils/tool-status.ts @@ -0,0 +1,15 @@ +// SPDX-License-Identifier: AGPL-3.0-only +// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +/** Mirrors NUDGE_TOOL_CALLS_STATUS in backend core/inference/tool_call_parser.py; keep in sync. */ +export const NUDGE_TOOL_CALLS_STATUS = "Nudging tool calls"; + +export type ToolStatusKind = "nudge" | "terminal" | "web"; + +/** Which glyph the badge shows: exact match for the nudge, "Running" prefix for sandbox tools, globe otherwise. */ +export function toolStatusKind(status: string): ToolStatusKind { + if (status === NUDGE_TOOL_CALLS_STATUS) { + return "nudge"; + } + return status.startsWith("Running") ? "terminal" : "web"; +} diff --git a/studio/frontend/tests/tool-status.test.ts b/studio/frontend/tests/tool-status.test.ts new file mode 100644 index 0000000000..b20585bd08 --- /dev/null +++ b/studio/frontend/tests/tool-status.test.ts @@ -0,0 +1,45 @@ +// SPDX-License-Identifier: AGPL-3.0-only +// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +import assert from "node:assert/strict"; +import test from "node:test"; + +import { + NUDGE_TOOL_CALLS_STATUS, + toolStatusKind, +} from "../src/features/chat/utils/tool-status.ts"; + +test("the nudge status is the exact string the backend sends", () => { + // Mirrors tool_call_parser.py, so a reword on either side must break here. + assert.equal(NUDGE_TOOL_CALLS_STATUS, "Nudging tool calls"); + assert.equal(toolStatusKind(NUDGE_TOOL_CALLS_STATUS), "nudge"); +}); + +test("sandbox tools keep the terminal glyph", () => { + for (const status of [ + "Running Python: print(1)", + "Running Python...", + "Running: ls -la", + "Running command...", + ]) { + assert.equal(toolStatusKind(status), "terminal", status); + } +}); + +test("every other status keeps the globe", () => { + for (const status of [ + "Searching: red square", + "Reading: unsloth.ai", + "Reading page...", + "Searching documents: quarterly report", + "Calling: get_weather", + ]) { + assert.equal(toolStatusKind(status), "web", status); + } +}); + +test("a status that merely mentions nudging is not the nudge itself", () => { + // Exact match only: a tool named after the phrase must not steal the spinner. + assert.equal(toolStatusKind("Calling: Nudging tool calls"), "web"); + assert.equal(toolStatusKind("Nudging tool calls again"), "web"); +}); From 9e2fc4985132473bbf914fdad3718141e35cf770 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:34:00 -0700 Subject: [PATCH 13/16] Studio: free the llama-server slot when a chat stream reaches [DONE] (#7564) * Studio: free the llama-server slot when a chat stream reaches [DONE] * Release the slot before yielding, only on a completed decode * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Tighten the comments added by this PR * Inline the done-sentinel check and use plain bools for the decode flags --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --- studio/backend/routes/inference.py | 41 ++- .../tests/test_gguf_stream_slot_release.py | 267 +++++++++++++++ .../test_gguf_stream_slot_release_ordering.py | 316 ++++++++++++++++++ 3 files changed, 622 insertions(+), 2 deletions(-) create mode 100644 studio/backend/tests/test_gguf_stream_slot_release.py create mode 100644 studio/backend/tests/test_gguf_stream_slot_release_ordering.py diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py index 12547277f5..d0a2d97f74 100644 --- a/studio/backend/routes/inference.py +++ b/studio/backend/routes/inference.py @@ -727,6 +727,7 @@ def _wants_stream_usage(payload) -> bool: _OPENAI_PASSTHROUGH_TERMINAL_GRACE_S = 2.0 _SSE_DONE_LINE = "data: [DONE]" +_SSE_DONE_CHUNK = "data: [DONE]\n\n" def _openai_passthrough_sse_line_terminal_state(raw_line: str) -> Optional[str]: @@ -2440,10 +2441,16 @@ async def _await_cancel_or_disconnect_then_close_client( return -async def _stop_local_disconnect_cancel_watcher(watcher) -> None: +async def _stop_local_disconnect_cancel_watcher(watcher, timeout_s: float = 5.0) -> None: + # Bounded: this runs in the stream's finally, so awaiting the watcher outright would let a + # wedged poll loop hold the response open forever. asyncio.wait neither cancels nor re-raises, + # and an abandoned watcher owns no resources. watcher.cancel() + done, _pending = await asyncio.wait({watcher}, timeout = timeout_s) + if not done: + return try: - await watcher + watcher.result() except (asyncio.CancelledError, Exception): pass @@ -9449,12 +9456,15 @@ async def openai_chat_completions( raise _openai_admission_http_exception(exc, status_code = 429) _tool_sentinel = object() + # True only once the sync generator returned on its own; see _gguf_decode_finished. + _tool_decode_finished = False _cancel_keys = (payload.cancel_id, payload.session_id, completion_id) _tracker = _TrackedCancel.for_payload(cancel_event, payload, *_cancel_keys) _tracker.__enter__() async def gguf_tool_stream(): + nonlocal _tool_decode_finished gen = None next_task = None stream_completed = False @@ -9542,6 +9552,7 @@ async def openai_chat_completions( if next_task.done(): next_task = None if event is _tool_sentinel: + _tool_decode_finished = True break # Anything after the gated tool_start means the user answered. @@ -9758,6 +9769,13 @@ async def openai_chat_completions( stream_started = True try: async for chunk in iterator: + # Release before the yield; see gguf_stream_chunks. + if ( + lease is not None + and _tool_decode_finished + and chunk == _SSE_DONE_CHUNK + ): + lease.release() yield chunk except asyncio.CancelledError: stream_cancelled = True @@ -10060,6 +10078,9 @@ async def openai_chat_completions( ) _gguf_sentinel = object() + # True only once the sync generator returned on its own: only then has _open_stream's + # client exited. A cancel still emits [DONE] without it. + _gguf_decode_finished = False if payload.stream: if _wants_multiple_choices(payload): @@ -10086,6 +10107,7 @@ async def openai_chat_completions( raise _openai_admission_http_exception(exc, status_code = 429) async def gguf_stream_chunks(): + nonlocal _gguf_decode_finished disconnect_watcher = asyncio.create_task( _await_disconnect_then_cancel(request, cancel_event) ) @@ -10130,6 +10152,7 @@ async def openai_chat_completions( if next_task.done(): next_task = None if cumulative is _gguf_sentinel: + _gguf_decode_finished = True break # Capture server metadata for the final usage chunk if isinstance(cumulative, dict): @@ -10292,6 +10315,20 @@ async def openai_chat_completions( stream_started = True try: async for chunk in iterator: + # The slot is idle once the sync generator returned and the stream ends + # with the plain sentinel. The finally only runs at ASGI teardown, so + # waiting for it starves the next request. Release before the yield: a + # stalled send() or a consumer that stops pulling parks us there, and + # Starlette never aclose()s a body iterator. Release is idempotent, so + # the finally stays the backstop. Exact equality, not endswith: + # _openai_stream_error_sse ends in the same sentinel before its + # cleanup runs, and that stream still owns the slot. + if ( + lease is not None + and _gguf_decode_finished + and chunk == _SSE_DONE_CHUNK + ): + lease.release() yield chunk except asyncio.CancelledError: stream_cancelled = True diff --git a/studio/backend/tests/test_gguf_stream_slot_release.py b/studio/backend/tests/test_gguf_stream_slot_release.py new file mode 100644 index 0000000000..4390f364c8 --- /dev/null +++ b/studio/backend/tests/test_gguf_stream_slot_release.py @@ -0,0 +1,267 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. + +"""A finished GGUF chat stream must free its llama-server slot at [DONE]. + +llama-server has a fixed slot count, gated by an admission lease. Releasing that lease only in +the stream's outer finally, which runs at ASGI teardown, let a wedged teardown pin a slot +llama-server had already freed, so the next chat request queued behind a finished generation +with no timeout to bound the wait. + +The wedge below stands in for the real one: the frontend never cancels its reader after [DONE] +(chat-api.ts), and uvicorn advertises ASGI spec_version 2.3, so Starlette's +OSError/ClientDisconnect path, the only disconnect detector _SameTaskStreamingResponse keeps, +cannot fire. +""" + +import asyncio +import json + +import pytest +from fastapi import FastAPI + +from auth.authentication import get_current_subject +from core.inference import llama_admission +import routes.inference as inference_route + + +@pytest.fixture(autouse = True) +def _fresh_queues(): + llama_admission.reset_llama_admission_queues() + yield + llama_admission.reset_llama_admission_queues() + + +def _active_slots() -> int: + with llama_admission._QUEUES_LOCK: + queues = list(llama_admission._QUEUES.values()) + return sum(queue.snapshot().active for queue in queues) + + +_ONE_SLOT = llama_admission.LlamaAdmissionConfig(max_queue = 4) + + +def _reserve_one_slot(): + """Take the single slot of a 1-parallel backend. Needs a running loop.""" + queue = llama_admission.get_llama_admission_queue("http://llama.test") + reservation = queue.reserve(capacity = 1, config = _ONE_SLOT) + return queue, reservation.lease_nowait() + + +def test_slot_is_freed_at_done_even_if_teardown_never_finishes(): + """Yield chunks, then wedge in the finally: without the release at [DONE] the slot stays + held for as long as the teardown is stuck, which is what starved the next request in CI. + """ + wedged = asyncio.Event() + + async def _stream(): + try: + yield 'data: {"choices": [{"delta": {"content": "hi"}}]}\n\n' + yield "data: [DONE]\n\n" + finally: + # Stand-in for a teardown that never completes. + await wedged.wait() + + async def _admitted(held): + iterator = _stream() + try: + async for chunk in iterator: + yield chunk + if held is not None and chunk == inference_route._SSE_DONE_CHUNK: + held.release() + finally: + if held is not None: + held.release() + + async def _drive(): + queue, lease = _reserve_one_slot() + assert lease is not None + assert _active_slots() == 1 + + seen = [] + saw_done = asyncio.Event() + + async def _consume(): + # Like Starlette's stream_response: it keeps pulling after the last chunk, so the + # generator resumes past [DONE] and only then runs into the wedged teardown. + async for chunk in _admitted(lease): + seen.append(chunk) + if chunk == inference_route._SSE_DONE_CHUNK: + saw_done.set() + + task = asyncio.create_task(_consume()) + try: + await asyncio.wait_for(saw_done.wait(), timeout = 5.0) + # Give the generator a turn to resume past the [DONE] yield and reach the wedge. + for _ in range(50): + if _active_slots() == 0: + break + await asyncio.sleep(0.01) + assert not task.done(), "teardown should still be wedged" + assert _active_slots() == 0, ( + "slot still held after [DONE]; the next chat request would " + "queue behind a generation that already finished" + ) + # A second caller must be admitted right away. + second = queue.reserve(capacity = 1, config = _ONE_SLOT).lease_nowait() + assert second is not None, "next request was refused a free slot" + second.release() + finally: + wedged.set() + task.cancel() + await asyncio.gather(task, return_exceptions = True) + return seen + + seen = asyncio.run(_drive()) + assert seen[-1] == "data: [DONE]\n\n" + + +def test_release_is_idempotent_so_the_finally_stays_a_backstop(): + async def _drive(): + _queue, lease = _reserve_one_slot() + assert _active_slots() == 1 + lease.release() + lease.release() + assert _active_slots() == 0 + + asyncio.run(_drive()) + + +def test_stopping_the_disconnect_watcher_cannot_hang(): + """The watcher stop runs in the stream's finally; it must be bounded.""" + + async def _drive(): + started = asyncio.Event() + + release = asyncio.Event() + + async def _unstoppable(): + started.set() + while not release.is_set(): + try: + await asyncio.sleep(0.01) + except asyncio.CancelledError: + # Swallow cancellation, as the real watcher does on its way out. + if release.is_set(): + raise + continue + + watcher = asyncio.create_task(_unstoppable()) + await started.wait() + # Would hang forever if the stop awaited the watcher outright. + await asyncio.wait_for( + inference_route._stop_local_disconnect_cancel_watcher(watcher, timeout_s = 0.2), + timeout = 5.0, + ) + assert not watcher.done(), "watcher should have been abandoned, not awaited" + release.set() + watcher.cancel() + await asyncio.gather(watcher, return_exceptions = True) + + asyncio.run(_drive()) + + +class _OneSlotGgufBackend: + """A loaded 1-parallel GGUF backend, the shape CI runs.""" + + is_loaded = True + model_identifier = "test/model.gguf" + base_url = "http://llama.test" + effective_parallel_slots = 1 + _is_audio = False + is_vision = False + supports_tools = False + + def generate_chat_completion(self, **kwargs): + yield "hi" + yield { + "type": "metadata", + "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}, + "timings": {"prompt_n": 3, "predicted_n": 1}, + "finish_reason": "stop", + } + + +def test_real_stream_frees_the_slot_at_done_with_a_wedged_teardown(monkeypatch): + """Drive the real ASGI route, wedged exactly where CI wedged. + + Hanging ``_stop_local_disconnect_cancel_watcher``, which runs in ``gguf_stream_chunks``'s + success-path finally, leaves a response that has sent [DONE] but cannot finish. + """ + monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: _OneSlotGgufBackend()) + monkeypatch.setattr(inference_route, "_effective_enable_tools", lambda payload: False) + + app = FastAPI() + app.include_router(inference_route.router) + app.dependency_overrides[get_current_subject] = lambda: "test-user" + + async def _drive(): + wedged = asyncio.Event() + + async def _hang(watcher, *args, **kwargs): + watcher.cancel() + await wedged.wait() + + monkeypatch.setattr(inference_route, "_stop_local_disconnect_cancel_watcher", _hang) + + body = json.dumps( + {"messages": [{"role": "user", "content": "hi"}], "stream": True} + ).encode() + scope = { + "type": "http", + "asgi": {"version": "3.0", "spec_version": "2.3"}, + "http_version": "1.1", + "method": "POST", + "scheme": "http", + "path": "/chat/completions", + "raw_path": b"/chat/completions", + "query_string": b"", + "root_path": "", + "headers": [ + (b"host", b"testserver"), + (b"content-type", b"application/json"), + (b"content-length", str(len(body)).encode()), + ], + "client": ("127.0.0.1", 12345), + "server": ("testserver", 80), + "app": app, + } + + sent_body = asyncio.Event() + frames = [] + + async def receive(): + if not frames: + return {"type": "http.request", "body": body, "more_body": False} + # Never disconnect: the browser keeps the socket open after [DONE]. + await asyncio.Event().wait() + + async def send(message): + frames.append(message) + if message.get("type") == "http.response.body": + chunk = message.get("body", b"").decode() + if chunk == inference_route._SSE_DONE_CHUNK: + sent_body.set() + + task = asyncio.create_task(app(scope, receive, send)) + try: + await asyncio.wait_for(sent_body.wait(), timeout = 20.0) + for _ in range(200): + if _active_slots() == 0: + break + await asyncio.sleep(0.01) + assert not task.done(), "response should still be wedged in teardown" + assert _active_slots() == 0, ( + "slot still held after [DONE] on the real route; the next chat " + "request would queue behind a finished generation" + ) + queue = llama_admission.get_llama_admission_queue("http://llama.test") + second = queue.reserve(capacity = 1, config = _ONE_SLOT).lease_nowait() + assert second is not None, "next request was refused a free slot" + second.release() + finally: + wedged.set() + task.cancel() + await asyncio.gather(task, return_exceptions = True) + + asyncio.run(_drive()) diff --git a/studio/backend/tests/test_gguf_stream_slot_release_ordering.py b/studio/backend/tests/test_gguf_stream_slot_release_ordering.py new file mode 100644 index 0000000000..7a8ceb4f53 --- /dev/null +++ b/studio/backend/tests/test_gguf_stream_slot_release_ordering.py @@ -0,0 +1,316 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. + +"""Ordering rules for the early admission release at ``data: [DONE]``. + +Freeing the llama-server slot at the sentinel is only correct when two things hold, and on a +one-slot backend both are load-bearing: + +1. The release happens *before* the sentinel reaches the ASGI ``send()``. Starlette's + ``stream_response`` suspends the body iterator at its ``yield`` for the whole of + ``await send(...)``, and uvicorn's ``send()`` awaits ``flow.drain()`` on a write-paused + transport, so a client that stops reading parks the generator there indefinitely. Starlette + never ``aclose()``s a body iterator either, so that generator's ``finally`` is left to GC. + +2. The sentinel really means "llama-server is done with this request". Two other emitters end + in the same bytes: ``_openai_stream_error_sse``, yielded from inside the still-suspended + generator's ``except`` block, and the cancel path, which breaks the read loop while the sync + generator is still parked on a yield inside ``_open_stream``'s httpx client. +""" + +import asyncio +import json +import threading + +import pytest +from fastapi import FastAPI + +from auth.authentication import get_current_subject +from core.inference import llama_admission +import routes.inference as inference_route + + +@pytest.fixture(autouse = True) +def _fresh_queues(): + llama_admission.reset_llama_admission_queues() + yield + llama_admission.reset_llama_admission_queues() + + +def _active_slots() -> int: + with llama_admission._QUEUES_LOCK: + queues = list(llama_admission._QUEUES.values()) + return sum(queue.snapshot().active for queue in queues) + + +class _OneSlotBackend: + """A loaded 1-parallel GGUF backend, the shape CI runs.""" + + is_loaded = True + model_identifier = "test/model.gguf" + base_url = "http://llama.test" + effective_parallel_slots = 1 + _is_audio = False + is_vision = False + supports_tools = False + + def __init__(self): + self.closing = threading.Event() + self.finish_close = threading.Event() + self.closed = threading.Event() + self.cancel_event = None + + def generate_chat_completion(self, **kwargs): + raise NotImplementedError + + +class _CompletingBackend(_OneSlotBackend): + def generate_chat_completion(self, **kwargs): + yield "hi" + yield { + "type": "metadata", + "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}, + "timings": {"prompt_n": 3, "predicted_n": 1}, + "finish_reason": "stop", + } + + +class _FailsMidStreamBackend(_OneSlotBackend): + """Still decoding when the route's own chunk handling blows up. + + ``gen`` stays parked on its ``yield`` until the stream's ``finally`` closes it, and only + that close drops the httpx stream llama-server is writing to. + """ + + def generate_chat_completion(self, **kwargs): + try: + yield "a" + yield "ab" + yield "abc" + except GeneratorExit: + self.closing.set() + # Stand in for the time llama-server needs to notice the drop and free its slot. + self.finish_close.wait(10.0) + self.closed.set() + raise + + +class _CancelledMidStreamBackend(_OneSlotBackend): + """Cancelled by the user halfway through, the Stop-button path.""" + + def generate_chat_completion( + self, + cancel_event = None, + **kwargs, + ): + self.cancel_event = cancel_event + try: + yield "a" + cancel_event.set() + yield "ab" + yield "abc" + except GeneratorExit: + self.closed.set() + raise + + +def _scope(app, body: bytes) -> dict: + return { + "type": "http", + "asgi": {"version": "3.0", "spec_version": "2.3"}, + "http_version": "1.1", + "method": "POST", + "scheme": "http", + "path": "/chat/completions", + "raw_path": b"/chat/completions", + "query_string": b"", + "root_path": "", + "headers": [ + (b"host", b"testserver"), + (b"content-type", b"application/json"), + (b"content-length", str(len(body)).encode()), + ], + "client": ("127.0.0.1", 12345), + "server": ("testserver", 80), + "app": app, + } + + +def _build_app(monkeypatch, backend): + monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend) + monkeypatch.setattr(inference_route, "_effective_enable_tools", lambda payload: False) + app = FastAPI() + app.include_router(inference_route.router) + app.dependency_overrides[get_current_subject] = lambda: "test-user" + return app + + +def _request_body() -> bytes: + return json.dumps({"messages": [{"role": "user", "content": "hi"}], "stream": True}).encode() + + +def test_slot_is_free_before_the_done_frame_reaches_send(monkeypatch): + """The release must not sit behind ``await send(...)``. + + uvicorn's ``send()`` awaits ``flow.drain()`` on a write-paused socket (h11_impl.py), so a + client that stops reading parks the body iterator on its ``yield`` indefinitely. Anything + after that ``yield`` is unreachable, and Starlette never ``aclose()``s the iterator, so the + outer ``finally`` is left to GC. + """ + backend = _CompletingBackend() + app = _build_app(monkeypatch, backend) + + async def _drive(): + body = _request_body() + frames = [] + slots_at_done = [] + finished = asyncio.Event() + + async def receive(): + if not frames: + return {"type": "http.request", "body": body, "more_body": False} + await asyncio.Event().wait() + + async def send(message): + frames.append(message) + if message.get("type") != "http.response.body": + return + if message.get("body", b"").decode() == "data: [DONE]\n\n": + # Sampled exactly where a stalled client would wedge. + slots_at_done.append(_active_slots()) + finished.set() + + task = asyncio.create_task(app(_scope(app, body), receive, send)) + try: + await asyncio.wait_for(finished.wait(), timeout = 20.0) + finally: + task.cancel() + await asyncio.gather(task, return_exceptions = True) + + assert slots_at_done == [0], ( + "the slot was still held while the [DONE] frame was being written; " + "a client that stops reading would pin it there indefinitely" + ) + + asyncio.run(_drive()) + + +def test_error_sentinel_keeps_the_slot_until_the_generator_is_closed(monkeypatch): + """``_openai_stream_error_sse`` ends in ``data: [DONE]`` but is not a finish. + + It is yielded from inside ``gguf_stream_chunks``'s ``except`` block, so the generator has + not yet run its ``finally``: the worker is undrained and ``gen`` is still open with + llama-server streaming into it. Freeing the slot there puts two callers on a one-slot + backend. + """ + backend = _FailsMidStreamBackend() + app = _build_app(monkeypatch, backend) + + calls = {"n": 0} + + def _boom(monitor_id, text): + calls["n"] += 1 + if calls["n"] >= 2: + raise RuntimeError("chunk handling failed") + + monkeypatch.setattr(inference_route.api_monitor, "append_reply", _boom) + + async def _drive(): + body = _request_body() + frames = [] + saw_error = asyncio.Event() + + async def receive(): + if not frames: + return {"type": "http.request", "body": body, "more_body": False} + await asyncio.Event().wait() + + async def send(message): + frames.append(message) + if message.get("type") != "http.response.body": + return + chunk = message.get("body", b"").decode() + # The error form: a payload line plus the sentinel, in one chunk. + if chunk.endswith("data: [DONE]\n\n") and chunk != "data: [DONE]\n\n": + saw_error.set() + + task = asyncio.create_task(app(_scope(app, body), receive, send)) + try: + await asyncio.wait_for(saw_error.wait(), timeout = 20.0) + # Wait until cleanup reaches gen.close(), so llama-server still holds the slot. + for _ in range(500): + if backend.closing.is_set(): + break + await asyncio.sleep(0.01) + assert backend.closing.is_set(), "cleanup never reached gen.close()" + assert _active_slots() == 1, ( + "slot handed out while the failed request still owned " + "llama-server; the next request would exceed the configured " + "parallelism" + ) + finally: + backend.finish_close.set() + task.cancel() + await asyncio.gather(task, return_exceptions = True) + + asyncio.run(_drive()) + + +def test_cancelled_stream_keeps_the_slot_until_the_generator_is_closed(monkeypatch): + """A cancelled stream emits the plain sentinel with ``gen`` still open. + + ``cancel_event.is_set()`` breaks the read loop at the top, so the sync generator never + reaches StopIteration and stays parked on a ``yield`` inside ``_open_stream``'s httpx + client. ``stream_completed`` is set all the same, which also makes the ``finally`` skip + ``gen.close()``, so ``data: [DONE]`` here does not mean llama-server is finished. + """ + backend = _CancelledMidStreamBackend() + app = _build_app(monkeypatch, backend) + + wedged = asyncio.Event() + + async def _hang(watcher, *args, **kwargs): + watcher.cancel() + await wedged.wait() + + monkeypatch.setattr(inference_route, "_stop_local_disconnect_cancel_watcher", _hang) + + async def _drive(): + body = _request_body() + frames = [] + saw_done = asyncio.Event() + + async def receive(): + if not frames: + return {"type": "http.request", "body": body, "more_body": False} + await asyncio.Event().wait() + + async def send(message): + frames.append(message) + if message.get("type") != "http.response.body": + return + if message.get("body", b"").decode() == "data: [DONE]\n\n": + saw_done.set() + + task = asyncio.create_task(app(_scope(app, body), receive, send)) + try: + await asyncio.wait_for(saw_done.wait(), timeout = 20.0) + for _ in range(50): + if _active_slots() == 0: + break + await asyncio.sleep(0.01) + assert backend.cancel_event is not None and backend.cancel_event.is_set() + assert ( + not backend.closed.is_set() + ), "test setup: the generator should still be open here" + assert _active_slots() == 1, ( + "slot freed on a cancelled stream whose llama-server request is " + "still open; the next request would exceed the configured " + "parallelism" + ) + finally: + wedged.set() + task.cancel() + await asyncio.gather(task, return_exceptions = True) + + asyncio.run(_drive()) From df63522369e239d32ab9833337ac2d1bfb472f53 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:50:38 -0700 Subject: [PATCH 14/16] Installer: stop requiring a developer toolchain on the consumer path (#7547) * Installer: stop requiring a developer toolchain on the consumer path A brand new Mac cannot install Studio at all. install.sh gates on `xcode-select -p` and exits 1 with 'Xcode Command Line Tools are required', and Linux exits 1 on any non-apt distro over cmake/gcc/git/libcurl headers. Nothing under either gate needs a toolchain. uv is a prebuilt binary, CPython comes from uv's managed python-build-standalone, llama.cpp and whisper.cpp are prebuilt downloads, Node is a pinned nodejs.org archive, and triton is skipped on macOS. unslothai/llama.cpp b10107-mix-1911198 publishes macos-arm64, macos-x64, linux-x64 and linux-arm64 builds covering cpu, cuda12, cuda13, rocm and vulkan. PR #6617 already dropped the Homebrew/cmake stop on macOS for this reason and just left the CLT stop behind. macOS: warn and continue when the CLT are absent. Linux: only a download transport (curl or wget) is fatal; build tooling warns. Both keep a hard git requirement for --local, which installs unsloth-zoo from a git+https URL. Both gates move into functions so tests/sh can extract them. The old inline form could not be reached by the tests/sh convention, which is why this shipped broken and stayed broken. test_macos_clt_gate.sh (19 assertions) and test_linux_deps_gate.sh (25) cover the clean machine, the CLT-stub shape where /usr/bin/git exists but fails, the non-apt distro, and the --local paths. Writing the Linux test caught a latent bug: the gate trimmed its list with $(echo ... | sed ...), so on a minimal image without sed the substitution yields empty and it reports 'all system dependencies found' on a machine with none of them. Replaced with parameter expansion. Also caps av<16 in the single-env constraints. av 16+ ships no cp313 macOS arm64 wheel, and it is a C extension over FFmpeg, so uv would silently fall back to a source build needing both a compiler and FFmpeg headers. Verified on GitHub-hosted macOS runners with /var/db/xcode_select_link, /Library/Developer/CommandLineTools, /Applications/Xcode*.app and Homebrew moved aside. macos-14, macos-15 and macos-26 fail on main and install cleanly with this; the recorded tool-invocation trace for the whole install is a single `xcode-select -p`, so nothing compiled and nothing installed a toolchain. * Linux: auto-install git rather than dropping it, and skip triton kernels without it Making git optional on Linux was too broad. studio/backend/requirements/ triton-kernels.txt line 2 is a git+https URL, so step 6/14 died with 'Cannot find command git' and failed the whole setup on ubuntu2404-root, ubuntu2404-arm-root and fedora41, all of which had been passing. The claim that nothing on the consumer path needs git holds on macOS, where triton is skipped, but not here. install.sh now auto-installs git through apt with the other optional tooling, so Debian and Ubuntu are unchanged. The triton kernels step skips with a message when git is absent instead of failing: they are a training speedup, not a boot requirement, and a GGUF chat install has no use for them. Six more assertions pin both halves. * macOS Intel: skip the one package with no x86_64 wheel The Intel clean-machine leg installed with the toolchain masked, then died in studio setup: subprocess.CalledProcessError: Command '['cmake', ...]' returned non-zero ERROR: Failed building wheel for pytorch_tokenizers pytorch_tokenizers publishes wheels for macOS arm64, linux x86_64, linux aarch64 and windows, but none for macOS x86_64 at any Python version, so uv falls back to an sdist that shells out to cmake. Nothing passes --only-binary, so the compiler-free property was an assumption rather than a contract, and Intel is where it broke. Marked so it installs everywhere except Intel macOS. Apple Silicon is unaffected. * Stop the optional dep gate from aborting the install _smart_apt_install exits rather than returns, and `|| true` does not catch an exit, so a box missing cmake or git aborted at the gate added to let it continue. Verified in sh, dash and bash. Run it in a subshell and re-raise only code 2, the NEED_SUDO handshake install.rs answers with an elevation prompt. install.sh treats a present-but-broken git as missing, but the Python side tested only shutil.which, so it promised to skip the git+https triton requirement and then fetched it anyway. Same check on both sides now. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Never elevate for optional build tools Re-raising code 2 turned the optional set into a NEED_SUDO handshake, so a box missing cmake or gcc got the desktop's mandatory permission dialog, whose Cancel drops back to not-installed. That re-imposes through a prompt the build-tool requirement this gate removes, and none of those tools are needed to run. Suppress the handshake for optional callers; a required package still elevates. Verified in sh, dash and bash. Also advance the progress bar on the no-git triton skip, which otherwise ends at 14/15. * Tighten the comments on the dependency gate * Correct why the PyAV cap is needed 16.0.0 does ship cp313-cp313-macosx_14_0_arm64; the comment claimed no cp313 wheel exists. The actual reason is the deployment target: 15.1.0 is macosx_13_0 and 16+ is macosx_14_0, so the cap is what keeps macOS 13 off a source build. * Tighten the installer gate comments * Cap cryptography on x86_64 macOS so the consumer install needs no Rust cryptography 49.0.0 (2026-06-12) dropped the macosx_10_9_universal2 wheel and now ships macosx_11_0_arm64 only, so x86_64 macOS has no wheel and uv falls back to the sdist. That build calls maturin, which pulls Rust and then fails at 'linking with cc failed' on a clean Mac without the Xcode Command Line Tools. It surfaced in the clean-machine leg mac macos-15-intel / mask / file, several minutes into the studio dependency step, which is exactly the up-front toolchain requirement this branch removes. 48.0.1 is the newest release carrying a universal2 wheel, and its cp39-abi3 / cp311-abi3 tags cover the 3.12 and 3.13 interpreters the installer creates. The cap is marker-scoped to darwin + x86_64, so arm64 macOS and every other platform still resolve to the latest. Lift it when cryptography ships an x86_64-capable macOS wheel again. Resolution of studio/backend/requirements/studio.txt under this constraints file gives 48.0.1 on x86_64-apple-darwin and 49.0.0 on aarch64-apple-darwin and x86_64-unknown-linux-gnu, on both 3.12 and 3.13. * Correct the av note now that cryptography also compiles on macOS * Never escalate for optional apt packages outside Tauri mode The optional bypass sat inside the TAURI_MODE branch, so a plain curl | sh install on a non-root Debian or Ubuntu box still fell through to the escalation branch and showed the default-yes permission prompt for cmake, GCC and the libcurl headers. That is exactly the toolchain this change set declared unnecessary on the consumer path, so the prompt asked for a password to install packages nothing here uses, and a headless run failed the same way instead of falling through to prebuilt llama.cpp. Move the check above the mode split so optional callers return 2 in both modes. Required packages such as curl still escalate unchanged. --------- Co-authored-by: danielhanchen Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --- install.sh | 196 +++++++++++----- .../backend/requirements/extras-no-deps.txt | 4 +- .../requirements/single-env/constraints.txt | 17 ++ studio/install_python_stack.py | 49 +++- tests/sh/test_linux_deps_gate.sh | 210 ++++++++++++++++++ tests/sh/test_macos_clt_gate.sh | 165 ++++++++++++++ 6 files changed, 574 insertions(+), 67 deletions(-) create mode 100755 tests/sh/test_linux_deps_gate.sh create mode 100755 tests/sh/test_macos_clt_gate.sh diff --git a/install.sh b/install.sh index 72f2455277..fc9aa0a431 100755 --- a/install.sh +++ b/install.sh @@ -800,8 +800,17 @@ _smart_apt_install() { return 0 fi - # In Tauri mode, report needed packages and exit — Rust handles elevation + # Optional callers never elevate, in any mode: nothing on the consumer path + # builds anything, so neither the terminal sudo prompt below nor the Tauri + # NEED_SUDO dialog (whose Cancel leaves the user not installed) may gate the + # run over unused tools. The caller falls through to prebuilt llama.cpp. + # Required packages such as curl still escalate. + if [ "${_SMART_APT_OPTIONAL:-false}" = true ]; then + return 2 + fi + if [ "$TAURI_MODE" = true ]; then + # Report needed packages and exit — Rust handles elevation. tauri_log "NEED_SUDO" "$_STILL_MISSING" exit 2 fi @@ -1998,67 +2007,142 @@ _maybe_reroute_strixhalo_to_2404() { _maybe_reroute_strixhalo_to_2404 || true # ── Check system dependencies ── -# cmake/git are only needed to *build* llama.cpp from source. Unsloth downloads a -# prebuilt by default, and setup.sh self-skips the source build when they're -# absent -- so macOS doesn't block on cmake (requiring it would force a manual -# Homebrew install). Linux keeps requiring them; its package manager has them. tauri_log "STEP" "Checking system dependencies" +# Without the Xcode CLT, macOS still ships /usr/bin/git as a stub that errors and pops +# a GUI dialog, so `command -v git` is not enough -- only running it tells the truth. +_has_working_git() { + command -v git >/dev/null 2>&1 || return 1 + git --version >/dev/null 2>&1 +} + +# macOS system-dependency check. A function so tests/sh can sed-extract it; the old +# inline form was untestable, which is why this gate shipped broken. +# +# The consumer install needs no developer toolchain: uv is a prebuilt binary, CPython +# is uv-managed, llama.cpp/whisper.cpp/Node are prebuilt downloads, and triton is +# skipped on macOS. Only `--local` needs git, for the unsloth-zoo git+https URL. +_check_macos_deps() { + _clt_missing=false + xcode-select -p >/dev/null 2>&1 || _clt_missing=true + + if [ "$STUDIO_LOCAL_INSTALL" = true ] && ! _has_working_git; then + echo "" + step "deps" "git is required for --local installs" "$C_ERR" + substep "--local installs unsloth-zoo from git+https://github.com/unslothai/unsloth-zoo," + substep "which needs a working git. Install the Xcode Command Line Tools:" + substep " xcode-select --install" + substep "Then re-run this script. A normal (non---local) install needs no compiler" + substep "and no git -- it uses prebuilt binaries and wheels only." + tauri_log "NEED_XCODE_CLT" "git" + return 1 + fi + + if [ "$_clt_missing" = true ]; then + # Not fatal, and no GUI dialog: firing xcode-select --install and exiting is + # what stranded clean Macs. + step "deps" "no Xcode Command Line Tools (not required)" "$C_WARN" + substep "Unsloth installs prebuilt binaries and wheels, so no compiler is needed." + substep "Install them only for a llama.cpp source build: xcode-select --install" + elif command -v cmake >/dev/null 2>&1; then + step "deps" "all system dependencies found" + else + # cmake is only for a source build, so its absence is not fatal. + step "deps" "using prebuilt llama.cpp (cmake not found)" "$C_WARN" + substep "Install cmake only if you want a source build: brew install cmake" + fi + return 0 +} + +# Linux/WSL system-dependency check. Same split as macOS, and a function for the same +# reason: tests/sh can extract it. +# +# Only a download transport is required. cmake, gcc and the libcurl headers exist +# solely for a llama.cpp source build the consumer path never does -- unslothai/ +# llama.cpp publishes linux-x64/arm64 prebuilts for cpu, cuda12, cuda13, rocm and +# vulkan. Requiring them turned every non-apt distro into a hard exit 1 over unused +# tooling. git follows macOS: --local only. +_check_linux_deps() { + _transport_missing=false + if ! command -v curl >/dev/null 2>&1 && ! command -v wget >/dev/null 2>&1; then + _transport_missing=true + fi + + # Wanted, never required: git fetches the triton_kernels git+https requirement (a + # training speedup), the rest serve the optional source build. Warn, never stop. + _optional_missing="" + command -v cmake >/dev/null 2>&1 || _optional_missing="$_optional_missing cmake" + _has_working_git || _optional_missing="$_optional_missing git" + command -v gcc >/dev/null 2>&1 || _optional_missing="$_optional_missing build-essential" + command -v curl-config >/dev/null 2>&1 || _optional_missing="$_optional_missing libcurl4-openssl-dev" + # Parameter expansion, not `sed`: sed may be absent on a minimal image, and a + # failed `$(... | sed ...)` yields "" -- "all found" on a machine that has none. + _optional_missing="${_optional_missing# }" + + if [ "$STUDIO_LOCAL_INSTALL" = true ] && ! _has_working_git; then + echo "" + step "deps" "git is required for --local installs" "$C_ERR" + substep "--local installs unsloth-zoo from git+https://github.com/unslothai/unsloth-zoo," + substep "which needs git. Install it with your package manager, then re-run." + substep "A normal (non---local) install needs no git and no compiler." + return 1 + fi + + # The one fatal case: nothing can be downloaded. apt is the only distro family we + # can drive unattended. + if [ "$_transport_missing" = true ]; then + if command -v apt-get >/dev/null 2>&1; then + echo "" + step "deps" "missing: curl" "$C_WARN" + substep "Needed to download uv, Python and the prebuilt inference engine." + _smart_apt_install curl + echo "" + else + echo "" + step "deps" "missing: curl (or wget)" "$C_ERR" + substep "Unsloth needs one of them to download uv, Python and the prebuilt" + substep "inference engine. Install one, then re-run setup:" + substep " Fedora/RHEL: sudo dnf install curl" + substep " Arch: sudo pacman -S --needed curl" + substep " openSUSE: sudo zypper install curl" + return 1 + fi + fi + + # Try apt for the optional set too; failing only costs the features warned about + # below. + if [ -n "$_optional_missing" ] && command -v apt-get >/dev/null 2>&1; then + step "deps" "installing optional build tools: $_optional_missing" "$C_DIM" + # Subshell because _smart_apt_install exits rather than returns, so `|| true` + # alone would not catch it. _SMART_APT_OPTIONAL suppresses every escalation + # path, so no install hinges on a prompt for tools nothing here needs. + ( _SMART_APT_OPTIONAL=true; _smart_apt_install $_optional_missing ) || true + _optional_missing="" + command -v cmake >/dev/null 2>&1 || _optional_missing="$_optional_missing cmake" + _has_working_git || _optional_missing="$_optional_missing git" + command -v gcc >/dev/null 2>&1 || _optional_missing="$_optional_missing build-essential" + command -v curl-config >/dev/null 2>&1 || _optional_missing="$_optional_missing libcurl4-openssl-dev" + _optional_missing="${_optional_missing# }" + fi + + if [ -n "$_optional_missing" ]; then + step "deps" "using prebuilt llama.cpp (missing: $_optional_missing)" "$C_WARN" + substep "Not required to run: Unsloth downloads a prebuilt inference engine." + case " $_optional_missing " in + *" git "*) substep "Without git the triton kernels training speedup is skipped." ;; + esac + else + step "deps" "all system dependencies found" + fi + return 0 +} + case "$OS" in macos) - # Xcode Command Line Tools provide the C/C++ compiler and git. - if ! xcode-select -p >/dev/null 2>&1; then - echo "" - echo "==> Xcode Command Line Tools are required." - echo " Installing (a system dialog will appear)..." - xcode-select --install /dev/null || true - echo " After the installation completes, please re-run this script." - exit 1 - fi - # cmake is only needed for a source build; the default prebuilt path - # doesn't use it, so its absence is not fatal -- no Homebrew prerequisite. - if command -v cmake >/dev/null 2>&1; then - step "deps" "all system dependencies found" - else - step "deps" "using prebuilt llama.cpp (cmake not found)" "$C_WARN" - substep "Install cmake only if you want a source build: brew install cmake" - fi + _check_macos_deps || exit 1 ;; linux|wsl) - MISSING="" - command -v cmake >/dev/null 2>&1 || MISSING="$MISSING cmake" - command -v git >/dev/null 2>&1 || MISSING="$MISSING git" - # curl or wget is needed for downloads; check both - if ! command -v curl >/dev/null 2>&1 && ! command -v wget >/dev/null 2>&1; then - MISSING="$MISSING curl" - fi - command -v gcc >/dev/null 2>&1 || MISSING="$MISSING build-essential" - # libcurl dev headers for llama.cpp HTTPS support - command -v curl-config >/dev/null 2>&1 || MISSING="$MISSING libcurl4-openssl-dev" - - MISSING=$(echo "$MISSING" | sed 's/^ *//') - if [ -n "$MISSING" ]; then - echo "" - step "deps" "missing: $MISSING" "$C_WARN" - substep "These are needed to build the GGUF inference engine." - if command -v apt-get >/dev/null 2>&1; then - _smart_apt_install $MISSING - else - echo " Automatic system package installation is supported on apt-based" - echo " Linux distributions (Ubuntu/Debian) only. Please install the" - echo " missing dependencies with your package manager, then re-run setup:" - echo " $MISSING" - echo "" - echo " Examples:" - echo " Fedora/RHEL: sudo dnf install cmake git gcc gcc-c++ make libcurl-devel" - echo " Arch: sudo pacman -S --needed cmake git base-devel curl" - echo " openSUSE: sudo zypper install cmake git gcc gcc-c++ make libcurl-devel" - exit 1 - fi - echo "" - else - step "deps" "all system dependencies found" - fi + _check_linux_deps || exit 1 ;; esac diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt index 3361af50dd..29d53ba204 100644 --- a/studio/backend/requirements/extras-no-deps.txt +++ b/studio/backend/requirements/extras-no-deps.txt @@ -15,7 +15,9 @@ trl==0.23.1 torch-c-dlpack-ext sentence_transformers==5.2.0 transformers==4.57.6 -pytorch_tokenizers +# No macOS x86_64 wheel at any version, so uv falls back to an sdist that shells out to +# cmake. Skipping it on Intel Macs keeps that install compiler-free. +pytorch_tokenizers; sys_platform != "darwin" or platform_machine == "arm64" kernels==0.12.1 # kernels<3.11 imports tomli as its tomllib fallback; --no-deps skips its own # marker dep, so list it here (no-op on the 3.12/3.13 default installs). diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt index 0a5619924a..7d3b9a081f 100644 --- a/studio/backend/requirements/single-env/constraints.txt +++ b/studio/backend/requirements/single-env/constraints.txt @@ -21,3 +21,20 @@ websockets>=15.0.1 anyio<4.14.0 pandas==2.3.3 + +# av (PyAV) 16+ builds its macOS arm64 wheels against macosx_14_0, so on macOS 13 none +# are installable and the resolver falls back to a source build, which needs FFmpeg +# headers the Xcode CLT do not supply and so fails however that Mac is equipped. +# 15.1.0 is the newest release with a macosx_13_0 arm64 wheel; 17+ moves to cp311-abi3 +# at macosx_14_0 too. +# +# The remaining sdist-only macOS defaults are pure Python, hence allowlisted in +# .github/scripts/clean-machine-assert.sh instead; cryptography below is the one +# other package that would compile. +av<16 + +# cryptography 49.0.0 dropped the macosx_10_9_universal2 wheel for arm64-only, so +# x86_64 macOS has no wheel and builds the sdist, needing Rust plus a working +# linker. 48.0.1 is the newest release with a universal2 wheel. Lift when +# cryptography ships an x86_64-capable macOS wheel again. +cryptography<49; sys_platform == "darwin" and platform_machine == "x86_64" diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py index 3243089656..886abe218b 100644 --- a/studio/install_python_stack.py +++ b/studio/install_python_stack.py @@ -2892,6 +2892,30 @@ def patch_package_file(package_name: str, relative_path: str, url: str) -> None: # -- Main install sequence --------------------------------------------- +def _has_working_git() -> bool: + """Match install.sh's _has_working_git: on PATH *and* actually runnable. + + A present-but-broken git (a bare xcrun shim) counts as missing there too. Testing + only shutil.which disagreed, so the installer promised to skip the git+https triton + requirement and then tried to fetch it anyway. + """ + exe = shutil.which("git") + if exe is None: + return False + try: + return ( + subprocess.run( + [exe, "--version"], + stdout = subprocess.DEVNULL, + stderr = subprocess.DEVNULL, + timeout = 30, + ).returncode + == 0 + ) + except (OSError, subprocess.SubprocessError): + return False + + def install_python_stack() -> int: global USE_UV, _STEP, _TOTAL _STEP = 0 @@ -3197,17 +3221,22 @@ def install_python_stack() -> int: _torchao_spec, ) - # 5. Triton kernels (no-deps, from source). Skip on Windows and macOS - # (no support). + # 5. Triton kernels (no-deps, from source). Skipped on Windows/macOS (no support) + # and without git (the requirement is a git+https URL); a training speedup + # only, so warn rather than fail the install. if not IS_WINDOWS and not IS_MACOS: - _progress("triton kernels") - pip_install( - "Installing triton kernels", - "--no-deps", - "--no-cache-dir", - req = REQ_ROOT / "triton-kernels.txt", - constrain = False, - ) + if not _has_working_git(): + _progress("triton kernels (skipped, no git)") + _safe_print(" no working git -- skipping triton kernels (training speedup only)") + else: + _progress("triton kernels") + pip_install( + "Installing triton kernels", + "--no-deps", + "--no-cache-dir", + req = REQ_ROOT / "triton-kernels.txt", + constrain = False, + ) if not IS_WINDOWS and not IS_MACOS and not NO_TORCH: _progress("flash-attn") diff --git a/tests/sh/test_linux_deps_gate.sh b/tests/sh/test_linux_deps_gate.sh new file mode 100755 index 0000000000..db25c5eb80 --- /dev/null +++ b/tests/sh/test_linux_deps_gate.sh @@ -0,0 +1,210 @@ +#!/bin/bash +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 +# +# Guards the Linux/WSL system-dependency gate in install.sh. +# +# History: the gate hard-required cmake, git, gcc and libcurl4-openssl-dev, installing +# them on apt distros and `exit 1`-ing everywhere else. Nothing on the consumer path +# builds anything, so it stranded every non-apt distro over unused tooling. +# +# The contract now: only a download transport (curl or wget) is fatal, build tooling +# is a warning, and git is required for --local only (unsloth-zoo git+https URL). +set -e + +SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" +INSTALL_SH="$SCRIPT_DIR/../../install.sh" +PASS=0 +FAIL=0 + +assert_contains() { + _label="$1"; _haystack="$2"; _needle="$3" + if echo "$_haystack" | grep -qF "$_needle"; then + echo " PASS: $_label" + PASS=$((PASS + 1)) + else + echo " FAIL: $_label (expected to find '$_needle')" + echo " ---- output ----"; echo "$_haystack" | sed 's/^/ | /' + FAIL=$((FAIL + 1)) + fi +} + +assert_not_contains() { + _label="$1"; _haystack="$2"; _needle="$3" + if echo "$_haystack" | grep -qF "$_needle"; then + echo " FAIL: $_label (found '$_needle' but should not)" + FAIL=$((FAIL + 1)) + else + echo " PASS: $_label" + PASS=$((PASS + 1)) + fi +} + +# ── Extract the functions under test ── +_FN_FILE=$(mktemp) +sed -n '/^_has_working_git()/,/^}/p' "$INSTALL_SH" > "$_FN_FILE" +sed -n '/^_check_linux_deps()/,/^}/p' "$INSTALL_SH" >> "$_FN_FILE" + +if ! grep -q '_check_linux_deps()' "$_FN_FILE"; then + echo "FAIL: could not extract _check_linux_deps from install.sh" + echo " (the gate must stay a top-level function so this test can reach it)" + exit 1 +fi + +_HARNESS=$(mktemp) +cat > "$_HARNESS" <<'HARNESS' +C_WARN=''; C_ERR=''; C_OK=''; C_DIM=''; C_RST='' +step() { echo "STEP $1 $2"; } +substep() { echo "SUBSTEP $1"; } +tauri_log() { echo "[TAURI:$1] $2"; } +# Records its args so a test can tell "asked apt for curl" from "asked for everything". +_smart_apt_install() { echo "APT_CALLED: $*"; } +HARNESS + +_BIN=$(mktemp -d) +_mk() { printf '#!/bin/sh\n%s\n' "$2" > "$_BIN/$1"; chmod +x "$_BIN/$1"; } + +# PATH is the sandbox and ONLY the sandbox, so unstocked tools are genuinely absent and +# the host's /usr/bin/cmake cannot leak in. bash must therefore be invoked absolutely. +_SH="${BASH:-/bin/bash}" + +_run_gate() { + # $1 = STUDIO_LOCAL_INSTALL + ( PATH="$_BIN"; export PATH + "$_SH" -c ". '$_HARNESS'; . '$_FN_FILE'; STUDIO_LOCAL_INSTALL=$1; _check_linux_deps; echo \"RC=\$?\"" 2>&1 ) +} + +echo "=== Fedora/Arch/openSUSE shape: curl present, no build tooling, no apt ===" +# Used to exit 1 with "supported on apt-based Linux distributions only". +rm -f "$_BIN"/* +_mk curl 'exit 0' +_out="$(_run_gate false)" +assert_contains "install proceeds" "$_out" "RC=0" +assert_contains "says the prebuilt is used" "$_out" "using prebuilt llama.cpp" +assert_contains "names what is missing" "$_out" "cmake" +assert_contains "says it is not required" "$_out" "Not required" +assert_not_contains "does not demand a package manager" "$_out" "apt-based" +assert_not_contains "does not reach apt for build tools" "$_out" "APT_CALLED" + +echo "=== wget instead of curl is an acceptable transport ===" +rm -f "$_BIN"/* +_mk wget 'exit 0' +_out="$(_run_gate false)" +assert_contains "install proceeds" "$_out" "RC=0" +assert_not_contains "does not ask apt for curl" "$_out" "APT_CALLED" + +echo "=== no transport at all, no apt: the one genuinely fatal case ===" +rm -f "$_BIN"/* +_out="$(_run_gate false)" +assert_contains "fails" "$_out" "RC=1" +assert_contains "names the missing transport" "$_out" "curl" +assert_contains "explains what it is needed for" "$_out" "download" +assert_contains "gives a non-apt remedy" "$_out" "dnf install curl" + +echo "=== no transport, apt available: auto-install curl and ONLY curl ===" +rm -f "$_BIN"/* +_mk apt-get 'exit 0' +_out="$(_run_gate false)" +assert_contains "install proceeds" "$_out" "RC=0" +assert_contains "asks apt for curl" "$_out" "APT_CALLED: curl" +assert_not_contains "does not ask apt for cmake" "$_out" "APT_CALLED: curl cmake" +# Build tooling still appears in the warning line, so match the apt call, not names. +assert_contains "apt asked for exactly curl" "$_out" "APT_CALLED: curl +" +assert_contains "build tooling only warned about" "$_out" "using prebuilt llama.cpp" + +echo "=== fully equipped machine: no warnings ===" +rm -f "$_BIN"/* +for t in curl cmake gcc curl-config git; do _mk "$t" 'exit 0'; done +_out="$(_run_gate false)" +assert_contains "install proceeds" "$_out" "RC=0" +assert_contains "reports everything found" "$_out" "all system dependencies found" +assert_not_contains "no prebuilt fallback warning" "$_out" "using prebuilt llama.cpp" + +echo "=== apt present: git is auto-installed, because triton_kernels needs it ===" +# Regression: making git optional without this failed at "6/14 triton kernels", whose +# requirement is a git+https URL. +rm -f "$_BIN"/* +_mk curl 'exit 0' +_mk apt-get 'exit 0' +_out="$(_run_gate false)" +assert_contains "install proceeds" "$_out" "RC=0" +assert_contains "apt is asked for git" "$_out" "git" +assert_contains "apt is actually called" "$_out" "APT_CALLED" + +echo "=== no apt and no git: warn about the triton skip, do not fail ===" +rm -f "$_BIN"/* +_mk curl 'exit 0' +_out="$(_run_gate false)" +assert_contains "install proceeds" "$_out" "RC=0" +assert_contains "names the consequence of no git" "$_out" "triton kernels" +assert_not_contains "does not call it required to run" "$_out" "is required" + +echo "=== --local without git: must fail loudly (matches macOS) ===" +rm -f "$_BIN"/* +_mk curl 'exit 0' +_out="$(_run_gate true)" +assert_contains "fails" "$_out" "RC=1" +assert_contains "explains why git is needed" "$_out" "unsloth-zoo" +assert_contains "says a normal install needs none" "$_out" "non---local" + +echo "=== --local with a git that exists but does not work ===" +# Mirrors the macOS CLT-stub shape: `command -v git` succeeds, running it fails. +rm -f "$_BIN"/* +_mk curl 'exit 0' +_mk git 'echo "broken" >&2; exit 1' +_out="$(_run_gate true)" +assert_contains "still fails" "$_out" "RC=1" + +echo "=== --local with a working git proceeds ===" +rm -f "$_BIN"/* +_mk curl 'exit 0' +_mk git 'exit 0' +_out="$(_run_gate true)" +assert_contains "install proceeds" "$_out" "RC=0" + +echo "=== optional apt packages never ask for elevation, in any mode ===" +# Regression: the optional bypass sat inside the TAURI_MODE branch, so a plain +# `curl | sh` on a non-root Debian box still hit the sudo prompt (default yes) and +# installed cmake, GCC and dev headers that nothing on the consumer path uses. +_APT_FN=$(mktemp) +{ + sed -n '/^_is_pkg_installed()/,/^}$/p' "$INSTALL_SH" + sed -n '/^_apt_distro_description()/,/^}$/p' "$INSTALL_SH" + sed -n '/^_can_read_tty()/,/^}$/p' "$INSTALL_SH" + sed -n '/^_smart_apt_install()/,/^}$/p' "$INSTALL_SH" +} > "$_APT_FN" + +_run_apt() { + # $1 = TAURI_MODE, $2 = _SMART_APT_OPTIONAL. apt-get always fails, as it does + # for a non-root user, so the function reaches its escalation decision. + rm -f "$_BIN"/* + _mk apt-get 'exit 100' + _mk sudo 'echo "ELEVATION_ATTEMPTED: $*"; exit 1' + ln -sf "$(command -v sed)" "$_BIN/sed" # the function trims its list with sed + # _APT_FN after _HARNESS so the real function replaces the recording stub. + ( PATH="$_BIN"; export PATH + "$_SH" -c ". '$_HARNESS'; . '$_APT_FN'; TAURI_MODE=$1; _SMART_APT_OPTIONAL=$2 + ( _smart_apt_install unsloth_absent_pkg ); echo \"RC=\$?\"" 2>&1 ) +} + +_out="$(_run_apt false true)" +assert_contains "optional: returns 2 so the caller can continue" "$_out" "RC=2" +assert_not_contains "optional: no sudo prompt" "$_out" "elevated permissions" +assert_not_contains "optional: sudo never invoked" "$_out" "ELEVATION_ATTEMPTED" + +_out="$(_run_apt true true)" +assert_contains "optional in Tauri: returns 2" "$_out" "RC=2" +assert_not_contains "optional in Tauri: no NEED_SUDO dialog" "$_out" "NEED_SUDO" + +_out="$(_run_apt false false)" +assert_contains "required: still escalates" "$_out" "ELEVATION_ATTEMPTED" + +_out="$(_run_apt true false)" +assert_contains "required in Tauri: still asks Rust to elevate" "$_out" "NEED_SUDO" + +rm -f "$_APT_FN" +rm -rf "$_BIN" "$_FN_FILE" "$_HARNESS" +echo "" +echo "=== $PASS passed, $FAIL failed ===" +[ "$FAIL" -eq 0 ] diff --git a/tests/sh/test_macos_clt_gate.sh b/tests/sh/test_macos_clt_gate.sh new file mode 100755 index 0000000000..2779df191e --- /dev/null +++ b/tests/sh/test_macos_clt_gate.sh @@ -0,0 +1,165 @@ +#!/bin/bash +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 +# +# Guards the macOS system-dependency gate in install.sh. +# +# History: the gate was inline top-level code running +# xcode-select -p || { xcode-select --install; exit 1; } +# so a brand-new Mac could not install at all, and being inline rather than a function +# it was out of reach of the tests/sh sed-extraction convention that would have caught +# it. +# +# The contract now: a consumer install must SUCCEED with no Xcode Command Line Tools +# (uv, CPython, llama.cpp/whisper.cpp/Node are all prebuilt, triton is skipped on +# macOS), while `--local` must still fail loudly: unsloth-zoo comes from a git+https +# URL. +set -e + +SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" +INSTALL_SH="$SCRIPT_DIR/../../install.sh" +PASS=0 +FAIL=0 + +assert_eq() { + _label="$1"; _expected="$2"; _actual="$3" + if [ "$_actual" = "$_expected" ]; then + echo " PASS: $_label" + PASS=$((PASS + 1)) + else + echo " FAIL: $_label (expected '$_expected', got '$_actual')" + FAIL=$((FAIL + 1)) + fi +} + +assert_contains() { + _label="$1"; _haystack="$2"; _needle="$3" + if echo "$_haystack" | grep -qF "$_needle"; then + echo " PASS: $_label" + PASS=$((PASS + 1)) + else + echo " FAIL: $_label (expected to find '$_needle')" + FAIL=$((FAIL + 1)) + fi +} + +assert_not_contains() { + _label="$1"; _haystack="$2"; _needle="$3" + if echo "$_haystack" | grep -qF "$_needle"; then + echo " FAIL: $_label (found '$_needle' but should not)" + FAIL=$((FAIL + 1)) + else + echo " PASS: $_label" + PASS=$((PASS + 1)) + fi +} + +# ── Extract the functions under test ── +_FN_FILE=$(mktemp) +sed -n '/^_has_working_git()/,/^}/p' "$INSTALL_SH" > "$_FN_FILE" +sed -n '/^_check_macos_deps()/,/^}/p' "$INSTALL_SH" >> "$_FN_FILE" + +if ! grep -q '_check_macos_deps()' "$_FN_FILE"; then + echo "FAIL: could not extract _check_macos_deps from install.sh" + echo " (the gate must stay a top-level function so this test can reach it)" + exit 1 +fi + +# Minimal harness: the output helpers install.sh would otherwise provide. +_HARNESS=$(mktemp) +cat > "$_HARNESS" <<'HARNESS' +C_WARN=''; C_ERR=''; C_OK=''; C_DIM=''; C_RST='' +step() { echo "STEP $1 $2"; } +substep() { echo "SUBSTEP $1"; } +tauri_log() { echo "[TAURI:$1] $2"; } +HARNESS + +_BIN=$(mktemp -d) + +# Each tool is absent, a working stub, or a broken stub mimicking the Xcode CLT shim +# (exists, exits non-zero). +_mk() { printf '#!/bin/sh\n%s\n' "$2" > "$_BIN/$1"; chmod +x "$_BIN/$1"; } + +# PATH is the sandbox and ONLY the sandbox, so unstocked tools are genuinely absent and +# the host's /usr/bin/git cannot leak in. bash must therefore be invoked absolutely. +_SH="${BASH:-/bin/bash}" + +_run_gate() { + # $1 = STUDIO_LOCAL_INSTALL + ( PATH="$_BIN"; export PATH + "$_SH" -c ". '$_HARNESS'; . '$_FN_FILE'; STUDIO_LOCAL_INSTALL=$1; _check_macos_deps; echo \"RC=\$?\"" 2>&1 ) +} + +echo "=== clean Mac: no CLT at all (xcode-select missing) ===" +rm -f "$_BIN"/* +_out="$(_run_gate false)" +assert_contains "does not exit 1" "$_out" "RC=0" +assert_contains "says CLT are not required" "$_out" "not required" +assert_not_contains "never claims CLT are required" "$_out" "are required" + +echo "=== clean Mac: CLT stubs present but non-functional (the real virgin-Mac shape) ===" +# With no CLT, /usr/bin/git EXISTS and fails when run, so `command -v git` succeeds. +# The gate must not be fooled by that. +rm -f "$_BIN"/* +_mk xcode-select 'exit 1' +_mk git 'echo "xcrun: error: invalid active developer path" >&2; exit 1' +_out="$(_run_gate false)" +assert_contains "consumer install proceeds" "$_out" "RC=0" +assert_contains "reports CLT absent but optional" "$_out" "not required" + +echo "=== --local with a non-functional git: must fail loudly ===" +_out="$(_run_gate true)" +assert_contains "fails" "$_out" "RC=1" +assert_contains "explains why git is needed" "$_out" "unsloth-zoo" +assert_contains "names the remedy" "$_out" "xcode-select --install" +assert_contains "emits a machine-readable marker" "$_out" "[TAURI:NEED_XCODE_CLT]" +assert_contains "says a normal install needs none" "$_out" "non---local" + +echo "=== --local with a working git: proceeds ===" +rm -f "$_BIN"/* +_mk xcode-select 'exit 1' +_mk git 'echo "git version 2.50.0"; exit 0' +_out="$(_run_gate true)" +assert_contains "--local proceeds when git works" "$_out" "RC=0" + +echo "=== CLT installed + cmake present ===" +rm -f "$_BIN"/* +_mk xcode-select 'echo /Library/Developer/CommandLineTools; exit 0' +_mk git 'echo "git version 2.50.0"; exit 0' +_mk cmake 'echo "cmake version 3.30.0"; exit 0' +_out="$(_run_gate false)" +assert_contains "all deps found" "$_out" "all system dependencies found" +assert_contains "rc 0" "$_out" "RC=0" + +echo "=== CLT installed, cmake missing: prebuilt path, not fatal ===" +rm -f "$_BIN"/* +_mk xcode-select 'echo /Library/Developer/CommandLineTools; exit 0' +_mk git 'echo "git version 2.50.0"; exit 0' +_out="$(_run_gate false)" +assert_contains "uses prebuilt llama.cpp" "$_out" "using prebuilt llama.cpp" +assert_contains "rc 0" "$_out" "RC=0" + +echo "=== the gate never fires the GUI installer on the consumer path ===" +# The dialog needs a GUI session a curl-piped or Tauri-spawned install does not have. +rm -f "$_BIN"/* +_mk xcode-select 'if [ "$1" = "--install" ]; then echo "GUI-DIALOG-FIRED"; fi; exit 1' +_out="$(_run_gate false)" +assert_not_contains "no GUI dialog on consumer path" "$_out" "GUI-DIALOG-FIRED" + +echo "=== _has_working_git distinguishes present-but-broken from working ===" +rm -f "$_BIN"/* +_mk git 'exit 1' +_r="$(PATH="$_BIN" "$_SH" -c ". '$_FN_FILE'; _has_working_git && echo yes || echo no")" +assert_eq "broken git stub -> no" "no" "$_r" +_mk git 'echo ok; exit 0' +_r="$(PATH="$_BIN" "$_SH" -c ". '$_FN_FILE'; _has_working_git && echo yes || echo no")" +assert_eq "working git -> yes" "yes" "$_r" +rm -f "$_BIN"/git +_r="$(PATH="$_BIN" "$_SH" -c ". '$_FN_FILE'; _has_working_git && echo yes || echo no")" +assert_eq "absent git -> no" "no" "$_r" + +rm -rf "$_BIN" "$_FN_FILE" "$_HARNESS" + +echo "" +echo "=== $PASS passed, $FAIL failed ===" +[ "$FAIL" -eq 0 ] || exit 1 From 4f0cbf0d81849b6e8c372f7144681f0a5ed285f6 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:52:00 -0700 Subject: [PATCH 15/16] Desktop: ask before quitting on top of a running install (#7550) * Desktop: ask before quitting on top of a running install This is the trigger neither #7492 nor #7490 addresses -- both start from a venv that is already broken. Confirmed: neither PR touches cleanup_child_processes. Quitting runs cleanup_child_processes -> install::stop_install, which SIGTERMs the installer's process group. In the reported session that landed at "5/10 studio deps", so the venv kept the CLI's dependencies and lost the server stack, and the next launch died on `import structlog`. Three minutes of installing, destroyed with no warning and no way back. So ask. Only from the tray Quit item -- a deliberate action with a UI present. The RunEvent::Exit path (OS shutdown, SIGTERM) is left alone: it must never block on a dialog nobody can answer. The call already runs off the menu callback thread, which is also what blocking_show requires. Closing the window was already safe (it hides to tray); this closes the remaining way to lose an install by accident. * Tighten comments in desktop quit-during-install guard * Condense comments in quit-during-install guard --------- Co-authored-by: danielhanchen --- studio/src-tauri/src/install.rs | 8 ++++++++ studio/src-tauri/src/main.rs | 30 ++++++++++++++++++++++++++++++ 2 files changed, 38 insertions(+) diff --git a/studio/src-tauri/src/install.rs b/studio/src-tauri/src/install.rs index d7226bf901..39d67dc427 100644 --- a/studio/src-tauri/src/install.rs +++ b/studio/src-tauri/src/install.rs @@ -783,6 +783,14 @@ pub fn record_install_intentional_stop(state: &InstallState, diagnostics: &Diagn } } +/// True while an installer runs; quitting now would leave a broken venv. +pub fn is_install_running(state: &InstallState) -> bool { + state + .lock() + .map(|install| install.child.is_some()) + .unwrap_or(false) +} + /// Stop a running install process gracefully. /// Unix: SIGTERM to process group -> wait up to 5s -> SIGKILL /// Windows: hidden taskkill /T /F to terminate the installer tree diff --git a/studio/src-tauri/src/main.rs b/studio/src-tauri/src/main.rs index a867700035..0d39217ecd 100644 --- a/studio/src-tauri/src/main.rs +++ b/studio/src-tauri/src/main.rs @@ -85,6 +85,33 @@ fn setup_custom_titlebar(app: &tauri::App) -> Result<(), Box bool { + use tauri_plugin_dialog::{DialogExt, MessageDialogButtons, MessageDialogKind}; + + let Some(install_state) = app.try_state::() else { + return true; + }; + if !install::is_install_running(&install_state) { + return true; + } + app.dialog() + .message( + "Unsloth Studio is still installing. Quitting now stops it part-way and \ + leaves the installation incomplete, so it will need to be repaired before \ + it can start.", + ) + .kind(MessageDialogKind::Warning) + .title("Installation in progress") + .buttons(MessageDialogButtons::OkCancelCustom( + "Quit anyway".to_string(), + "Keep installing".to_string(), + )) + .blocking_show() +} + fn cleanup_child_processes(app: &tauri::AppHandle) { let diagnostics_state = app .try_state::() @@ -138,6 +165,9 @@ fn setup_tray(app: &tauri::App) -> Result<(), Box> { // leaving the backend orphaned. let app_handle = app.clone(); std::thread::spawn(move || { + if !confirm_quit_during_install(&app_handle) { + return; + } cleanup_child_processes(&app_handle); app_handle.exit(0); }); From 00646632bcdf3ee56dd07fef6d6f5a624a50beec Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 28 Jul 2026 18:52:25 -0700 Subject: [PATCH 16/16] Tests: import bitsandbytes before the GPU-free harness spoofs CUDA (#7582) * Tests: import bitsandbytes before the GPU-free harness spoofs CUDA The CPU test harness patches torch.cuda.is_available to return True so device_type.py's cache captures "cuda" on a GPU-less runner. bitsandbytes reads the same flag at import time to decide whether to load its CUDA backend, and that backend reads torch._C._cuda_getCurrentRawStream, which a CPU-only torch build does not expose. An import landing inside the spoof window therefore raises, Python drops bitsandbytes from sys.modules while leaving its submodules cached, and every later import returns a module with no .functional, so unsloth/kernels/utils.py dies at module scope. Import bitsandbytes before the window so it stays on its CPU backend and remains fully usable, rather than being degraded to unavailable. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: danielhanchen Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --- tests/conftest.py | 27 ++++++ .../test_conftest_bitsandbytes_preimport.py | 85 +++++++++++++++++++ 2 files changed, 112 insertions(+) create mode 100644 tests/python/test_conftest_bitsandbytes_preimport.py diff --git a/tests/conftest.py b/tests/conftest.py index 3478a19af8..aaeeb840ce 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -123,7 +123,34 @@ def _install_device_type_stub(name: str) -> None: sys.modules[name] = stub +def _preimport_bitsandbytes() -> None: + """Bind bitsandbytes against the real torch before the CUDA spoof below. + + `bitsandbytes/__init__.py` runs `if torch.cuda.is_available(): from .backends.cuda + import ops`, and that module reads `torch._C._cuda_getCurrentRawStream`, which a + CPU-only torch build does not expose. `_preload_device_type` patches + `torch.cuda.is_available` to return True, so a bitsandbytes import landing inside + that window takes the CUDA branch and dies with AttributeError. + + Python then drops `bitsandbytes` from sys.modules but leaves `bitsandbytes.functional` + and the rest of its submodules cached, so the next import re-executes __init__ against + those cached submodules, re-binds nothing, and hands back a module with no + `.functional`. `unsloth/kernels/utils.py` reads `bnb.functional.get_ptr` at module + scope, so every later `import unsloth` in that process dies with + "module 'bitsandbytes' has no attribute 'functional'". + + Importing first, outside the window, keeps bitsandbytes on its CPU backend and fully + usable. Must stay ahead of the `_preload_device_type` calls below. + """ + try: + import bitsandbytes # noqa: F401 + except Exception: + # A genuinely absent or broken wheel is unsloth's own degradation path. + pass + + if not _has_real_accelerator(): + _preimport_bitsandbytes() if not _preload_device_type("unsloth_zoo", prereqs = ("utils",)): _install_device_type_stub("unsloth_zoo.device_type") if not _preload_device_type("unsloth"): diff --git a/tests/python/test_conftest_bitsandbytes_preimport.py b/tests/python/test_conftest_bitsandbytes_preimport.py new file mode 100644 index 0000000000..8ed80d6232 --- /dev/null +++ b/tests/python/test_conftest_bitsandbytes_preimport.py @@ -0,0 +1,85 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. + +"""Guard the ordering that keeps bitsandbytes usable under the GPU-free harness. + +tests/conftest.py patches `torch.cuda.is_available` to return True so +`device_type.py`'s @cache captures "cuda" on a GPU-less runner. bitsandbytes reads +that same flag at import time to decide whether to import its CUDA backend, and that +backend touches `torch._C._cuda_getCurrentRawStream`, absent from CPU-only torch +builds. A bitsandbytes import landing inside the spoof window therefore raises, and +the failure is not recoverable within the process: Python drops `bitsandbytes` from +sys.modules while leaving its submodules cached, so every later import returns a +module with no `.functional`, and `unsloth/kernels/utils.py` dies at module scope. + +Clearing sys.modules is not a way out either -- re-executing `bitsandbytes._ops` +raises "Tried to register an operator ... multiple times". The import simply must not +fail, which is what `_preimport_bitsandbytes()` guarantees by running first. + +Source-level rather than behavioural on purpose: the failure needs a CPU-only torch +build to reproduce, so a runtime assertion would pass vacuously wherever CUDA torch +is installed, which is most developer machines. +""" + +from __future__ import annotations + +import ast +from pathlib import Path + +CONFTEST = Path(__file__).resolve().parents[1] / "conftest.py" + + +def _accelerator_guard_body(tree: ast.Module) -> list[ast.stmt]: + for node in tree.body: + if isinstance(node, ast.If) and "_has_real_accelerator" in ast.dump(node.test): + return node.body + raise AssertionError("tests/conftest.py has no `if not _has_real_accelerator():` block") + + +def _called_names(body: list[ast.stmt]) -> list[str]: + names = [] + for stmt in body: + for node in ast.walk(stmt): + if isinstance(node, ast.Call) and isinstance(node.func, ast.Name): + names.append(node.func.id) + return names + + +def test_conftest_defines_the_bitsandbytes_preimport(): + tree = ast.parse(CONFTEST.read_text(encoding = "utf-8")) + defined = {n.name for n in tree.body if isinstance(n, ast.FunctionDef)} + assert "_preimport_bitsandbytes" in defined, ( + "tests/conftest.py must define _preimport_bitsandbytes(); without it a " + "bitsandbytes import inside the CUDA spoof window permanently breaks " + "`import unsloth` for the rest of the process" + ) + + +def test_bitsandbytes_is_preimported_before_the_cuda_spoof(): + tree = ast.parse(CONFTEST.read_text(encoding = "utf-8")) + called = _called_names(_accelerator_guard_body(tree)) + + assert "_preimport_bitsandbytes" in called, ( + "_preimport_bitsandbytes() is never called inside the " + "`if not _has_real_accelerator():` block" + ) + assert "_preload_device_type" in called, "conftest no longer calls _preload_device_type" + assert called.index("_preimport_bitsandbytes") < called.index("_preload_device_type"), ( + "_preimport_bitsandbytes() must run BEFORE _preload_device_type(), which is what " + "patches torch.cuda.is_available; importing bitsandbytes inside that window makes " + "it take its CUDA backend on a CPU-only torch and poisons sys.modules" + ) + + +def test_preimport_swallows_a_genuinely_missing_wheel(): + """An absent bitsandbytes stays unsloth's own degradation path, not a collection error.""" + tree = ast.parse(CONFTEST.read_text(encoding = "utf-8")) + fn = next( + n + for n in tree.body + if isinstance(n, ast.FunctionDef) and n.name == "_preimport_bitsandbytes" + ) + assert any(isinstance(node, ast.Try) for node in ast.walk(fn)), ( + "_preimport_bitsandbytes() must guard its import with try/except so a missing or " + "broken wheel does not turn into a collection error" + )