unsloth/studio/backend/core/training/worker.py
Daniel Han 80d5acafb4
studio: install flash-linear-attention and tilelang for Qwen3.5 family (#5434)
* studio: install flash-linear-attention and tilelang for Qwen3.5 family

Studio currently only installs causal-conv1d for qwen3.5 / qwen3.6 /
qwen3-next models. Without flash-linear-attention installed alongside
it, transformers' Qwen3.5 fast-path gate stays False and the model
falls back to a pure-PyTorch loop for the GatedDeltaNet layers. In a
60-step run on unsloth/Qwen3.5-2B on B200, this fallback costs ~2.35x
vs the full fast path.

On top of that, FLA dispatches its hottest GDN kernels through a
TileLang backend when tilelang is importable. Adding tilelang plus a
pinned apache-tvm-ffi gives another ~26% on the same workload (4.73
s/step to 3.50 s/step) and is what users have been getting indirectly
when they install mamba-ssm (mamba-ssm transitively pulls tilelang and
pins apache-tvm-ffi<=0.1.9, which is the last working version on
sm_100; 0.1.10 and 0.1.11 crash Triton with misaligned address).

Changes:
  * _ensure_flash_linear_attention: pure-Python PyPI install gated on
    the same model match set as _ensure_causal_conv1d_fast_path.
  * _ensure_tilelang_backend: installs apache-tvm-ffi==0.1.9 and
    tilelang==0.1.8 in one pip resolve so the tvm-ffi pin wins over
    tilelang's >=0.1.2 constraint. Gated on the Qwen3.5 family only;
    SSM models (Nemotron-H, Falcon-H1, Granite-H, LFM2) do not use
    FLA's GDN dispatch.
  * UNSLOTH_STUDIO_SKIP_TILELANG_INSTALL=1 escape hatch matching the
    flash-attn pattern.
  * Orchestration block reordered: causal-conv1d -> fla -> mamba-ssm
    -> tilelang -> flash-attn (long context).
  * 7 new tests covering the new helpers, including SSM-model skip,
    skip-env, full Qwen3 family name variants, and graceful pip
    install failure.

Combined Qwen3.5-2B-Vision step time on B200 in our bench goes from
5.0 s/step (current Studio: causal-conv1d only) to 3.5 s/step
(causal-conv1d + fla + tilelang), a 1.43x speedup with no notebook
or user code changes required.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* tests/studio: accept new grad_norm arg in MLX smoke _on_step callback

The MLX trainer's step callback now passes a ninth positional argument
(grad_norm) per unsloth_zoo/mlx/trainer.py's documented signature
``fn(step, total_steps, loss, lr, tokens_sec, peak_gb, elapsed,
num_tokens, grad_norm=None)``. The smoke's local ``_on_step`` was still
defined with eight, so every per-step invocation raised
``TypeError: _on_step() takes 8 positional arguments but 9 were given``,
``losses_per_step`` never got populated, and the post-train
``assert len(losses_per_step) == 7`` failed.

Add the ninth parameter with a default and surface the gradient norm in
the per-step log line when present.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* ci: retrigger after zoo drift + IPython fixes landed in main

* tests/studio: pin max_grad_value=0 in MLX smoke so max_grad_norm=1.0 wins

unsloth_zoo PR #5340 added per-element gradient clipping to MLXTrainer
and defaulted ``MLXTrainingConfig.max_grad_value = 5.0``. When both
``max_grad_norm`` and ``max_grad_value`` are set, the trainer warns:

  Unsloth: max_grad_norm and max_grad_value are both enabled;
  ignoring max_grad_norm in favor of max_grad_value.

and silently drops the test's ``max_grad_norm=1.0``. +-5.0 per-element
is far too loose for this 270M Gemma-3 LoRA r=8 (attention + MLP) at
bs=2 ga=3 lr=1e-3: the update direction is no longer norm-bounded, so
losses overshoot and the model fails to memorise the training row.

Reproduced on a CUDA mirror (scripts/cuda_mlx_mirror_sim.py):

  norm_1       (max_grad_norm=1.0, no clip): losses 7.64 -> 0.006,
                generation contains 'Unsloth' (the smoke's pass case)
  clip_value_5 (max_grad_norm=0, clip+-5.0): losses 7.29 -> 8.39
                (DIVERGED after step 4), generation gibberish, no
                'Unsloth' -- exactly the failure surfaced on PR 5434
                once the _on_step 9-arg fix let the smoke past the
                training loop.

Pin ``max_grad_value=0.0`` so the smoke uses the same ``max_grad_norm=
1.0`` clipping it was designed against. Leaves the new default in
place for everyone else; only the smoke needs deterministic clipping
to validate the round-trip.

* tests/studio: clarify why MLX smoke pins max_grad_value=0

Refresh the rationale comment to reflect the new default landing in
unslothai/unsloth-zoo#652 (max_grad_value=1.0, not 5.0). The smoke
still needs the explicit pin because neither default value reliably
converges in 7 steps at seed=3407:

  max_grad_value=5.0 -- diverges after step 4 (loss 7.3 -> 8.4)
  max_grad_value=1.0 -- stalls (loss ~3.2 plateau across seeds)
  max_grad_value=0.5/0.25/0.1 -- noisier still
  max_grad_norm=1.0  -- cleanly drops loss to <0.01, emits "Unsloth!"

Mention both the historical 5.0 default and the new 1.0 default in
the comment so future readers do not assume the smoke is dead code
referencing a removed knob, and point to the CUDA mirror scripts
(cuda_mlx_mirror_sim.py + cuda_mlx_clip1_vs_norm1.py) for the
empirical evidence.

No behaviour change; comment-only refresh.

* tests/studio: replace fragile substring gate with loss + round-trip gates

The MLX smoke's three "EXPECT in completion" assertions assume the
trained model will greedy-emit the exact "Unsloth" token after the
prompt. On MLX a single near-zero-loss adamw step at the smoke's
fixed seed=3407 can perturb the final-step logits enough that greedy
decoding picks a wrong first token even while the teacher-forced loss
on the training row stays essentially zero (the smoke captures this
exact state -- step 6 loss=0.049, step 7 grad=36.7, step 7 loss=0.17;
completion goes from "Unsloth!" to "5 lbs!"). Reproduced extensively
on CUDA via scripts/cuda_mlx_step7_*.py: at seed=3407 only one config
in a 9-cell sweep lands inside the "Unsloth"-emitting basin, and only
1/3 seeds at that config pass. This is a property of the assertion,
not of save/reload correctness.

Refactor the three assertions to gate on what the smoke is actually
trying to verify:

  in_memory:
    - hard gate: post_train_loss < 1.0 (training memorised the row).
    - soft check: log whether completion contains EXPECT_IN_OUTPUT
      into metrics["in_memory_generation_has_expected"]; print a
      WARN when missing instead of failing.

  lora / merged reload:
    - hard gate: reload output must equal the in-memory completion
      saved in train_metrics.json. This is the actual save/reload
      invariant -- the reloaded weights have to reproduce whatever
      the in-memory model produced. Falls back to the original
      gibberish gate if train_metrics.json is unavailable.

  gguf reload:
    - hard gate: llama.cpp produced usable, non-empty output after
      the prompt (>=4 chars). llama.cpp's tokenizer + sampling differ
      from mlx_lm so byte-exact match isn't sound. Log
      gguf_has_expected for visibility.

Result: the smoke still gates on the real failure modes (training
didn't memorise, save/reload corrupted weights, llama.cpp produced
no output), without depending on the brittle "Unsloth as first
greedy-decoded token" guarantee that MLX's step-7 numerics can break
without harming any save/reload semantics.

Cross-version constraint: no transformers / trl API touched.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* tests/studio: gate MLX reload on training-row loss, not greedy text

The strict reload assertion (out == in_mem_out) failed on macOS:
in-memory completion was '5 lbs!' and the reloaded completion was
'_________________________'. Both are corrupted by the same MLX
step-7 grad spike (see scripts/cuda_mlx_step7_*), but greedy decoding
can pick a different first token at near-zero teacher-forced loss
even when weights are byte-identical, so exact text equality is not
the right round-trip invariant.

Replace with teacher-forced loss equality on TRAIN_TEXT: the
reloaded model must reach essentially the same post_train_loss the
in-memory model recorded. That is the real save/reload correctness
gate, robust to MLX's near-zero-loss adamw greedy-decode
perturbation. Falls back to a non-empty-body check when
train_metrics.json is missing.

CUDA mirror at this seed converges cleanly to ~0.006 loss; on MLX
post_train_loss < 1.0 still holds via the existing memorisation
gate. The completion text and "matches in-memory" flag are still
recorded in metrics for visibility, just not gated on.

* ci: retrigger Backend CI after transient pwsh-startup timeout

* ci: retrigger MLX dispatch after pytorch CDN DNS flake

* studio: harden FLA + tilelang installers per reviewer feedback

Addresses bot review on #5434:

  * Narrow `_ensure_flash_linear_attention` from `_model_wants_causal_conv1d`
    (which also matches Nemotron-H / Falcon-H1 / Granite-H / LFM2) to
    `_model_wants_tilelang` (Qwen3.5 / Qwen3.6 / Qwen3-Next only). True
    SSM families take the mamba_ssm path and never call FLA's GDN
    kernels, so installing FLA there is wasted bandwidth.

  * Pin both `flash-linear-attention==0.5.0` and `fla-core==0.5.0` and
    install with `--no-deps`. Otherwise pip resolves fla-core's
    declared `torch>=2.7.0` requirement and may silently upgrade the
    Studio venv's torch on environments running torch 2.4/2.5/2.6.

  * Skip both installs on Python <3.10 (FLA, fla-core, and tilelang
    all declare `Requires-Python: >=3.10`). On older interpreters the
    pip install would fail every launch and leave the worker on the
    slow torch fallback while still claiming to have set up the fast
    path.

  * Skip tilelang install on non-Linux platforms. `tilelang==0.1.8`
    only publishes Linux x86_64 / aarch64 and macOS arm64 wheels.
    Falling back to its 93MB sdist on a Studio worker is undesirable.

  * Detect an existing `apache-tvm-ffi` 0.1.10 / 0.1.11 install and
    force a reinstall to 0.1.9 with `--force-reinstall --no-deps`.
    Previously the import-only probe returned early and left the
    broken version in place, which crashes Triton on sm_100.

  * Add a 600s timeout to the tilelang and FLA subprocess.run calls,
    matching the existing flash-attn install pattern, so a network
    hang cannot block the training subprocess indefinitely.

  * 13 new / updated tests covering all six guards plus the
    pinned-spec, timeout, and force-reinstall code paths.

Total: 21 passing tests (8 original + 13 new / updated).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: address reviewer.py P1/P2 findings on FLA + tilelang installers

Twelve-reviewer aggregated review on this PR flagged several real
correctness bugs in the first hardening pass. Fixes:

P1:
  * Add UNSLOTH_STUDIO_SKIP_FLA_INSTALL escape hatch for symmetry
    with UNSLOTH_STUDIO_SKIP_TILELANG_INSTALL and the existing
    UNSLOTH_STUDIO_SKIP_FLASHATTN_INSTALL.
  * Install einops alongside fla-core. `--no-deps` was suppressing
    fla-core's only non-torch runtime dep, so on a clean venv
    `import fla.modules` raised ModuleNotFoundError even though pip
    exited 0.
  * Drop --no-deps from the tilelang force-reinstall path. tilelang
    needs z3-solver, ml-dtypes, cloudpickle, etc. at runtime;
    --force-reinstall --no-deps left libz3.so missing and
    `import tilelang` raised OSError on the next training subprocess.
  * Skip FLA install when installed torch is below 2.7.0
    (fla-core declares torch>=2.7.0). Otherwise users on Studio's
    supported torch 2.4/2.5/2.6 stacks get an incompatible FLA
    installed silently.

P2:
  * Replace bare `except ImportError` probes with helpers that catch
    `Exception` so a broken native package (OSError on missing
    .so, RuntimeError in __init__, ...) does not kill the worker
    before the fallback path can run.
  * Tighten the tilelang platform guard from "any linux" to
    "linux + machine in {x86_64, aarch64, ...}" so ppc64le / s390x /
    armv7 do not fall through and download the 93 MB tilelang sdist.
  * Add --only-binary=:all: to the tilelang install command. The
    comment already said we never want the sdist; now the pip
    invocation enforces it.
  * Verify both FLA and tilelang are importable after pip exits 0;
    if not, report and continue on the fallback path.

6 new tests bring the suite to 27 passing (was 21).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: pin packaging + triton with FLA --no-deps install

An end-to-end install simulation in a fresh venv caught a real
regression: `fla/utils.py` does `from packaging import version` and
`import triton` at module load, but fla-core's METADATA only declares
einops + torch. With `--no-deps` the worker would land FLA in any
runtime that lacks packaging (e.g. minimal torch builds) and the
post-install import probe would fall back to the torch GDN loop
silently.

Add `packaging` and `triton` to `_FLA_RUNTIME_DEPS` so the install
spec list always carries them. Tests updated to assert both are now in
the install command.

* studio: hook transformers' fast-path gates for just-in-time FLA + causal-conv1d install

The substring-based detection in this PR (`_model_wants_tilelang` /
`_model_wants_causal_conv1d`) is brittle: it depends on what the user
typed for the model name, not on what the architecture actually needs.
Users typing custom model paths, future Qwen3.7 / non-Qwen GDN
architectures, and any model whose author renamed it would silently
fall back to the torch loop.

The correct signal is the one transformers itself uses to gate the
fast path. `transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py`
does at module import time:

    if is_causal_conv1d_available():
        from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
    if is_flash_linear_attention_available():
        from fla.modules import FusedRMSNormGated
        from fla.ops.gated_delta_rule import (
            chunk_gated_delta_rule, fused_recurrent_gated_delta_rule,
        )

Wrap both gates so the first call (always at modeling import, before
any forward pass) installs the matching kernel synchronously and
delegates to the original function. Any model whose architecture
queries those gates auto-triggers the install; models that never
query them (Llama, Gemma, dense Qwen, ...) never pay the cost.

Mechanics:

  - Split `_ensure_flash_linear_attention` and `_ensure_tilelang_backend`
    into `_unconditional` variants (no substring gate, retains python
    / torch / platform / skip-env guards) plus thin substring wrappers
    used by the legacy fallback path.
  - New `_install_fast_path_hooks(event_queue)` patches both gates on
    `transformers.utils.import_utils` AND sweeps `sys.modules` so any
    modeling file that already did `from ... import is_X` sees the
    wrapper (the local binding survives a module-level reassignment).
  - Wrappers clear the original's `lru_cache` before delegating, install
    on False, re-check, and short-circuit on subsequent calls.
  - Set `UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS=1` to fall back to the
    substring path.

Verified end-to-end against `transformers.models.qwen3_5_moe`:

  PRE_STATE fla=False tilelang=False causal_conv1d=False
  HOOK_INSTALLED
  Hook fired for is_causal_conv1d_available; installing kernel...
  Installing prebuilt causal-conv1d wheel...
  Hook fired for is_flash_linear_attention_available; installing kernel...
  Installing flash-linear-attention==0.5.0 (with fla-core==0.5.0) for the fast path...
  Installed flash-linear-attention for the FLA fast path
  Installing TileLang backend (apache-tvm-ffi==0.1.9, tilelang==0.1.8)...
  Installed TileLang backend for FLA fast path
  MODELING_IMPORT_OK
  FAST_PATH_SYMBOLS {"chunk_gated_delta_rule": true,
                     "fused_recurrent_gated_delta_rule": true,
                     "FusedRMSNormGated": true,
                     "causal_conv1d_fn": true,
                     "causal_conv1d_update": true}
  POST_STATE fla=True tilelang=True causal_conv1d=True

Adds 9 new tests covering: install-on-False, skip-on-True, idempotency,
install-failure handling, env-disable, lru_cache clear, sys.modules
rebind, missing-transformers fallback, substring fallback. Total
test count is now 36 (was 27).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: address reviewer.py n=12 findings on the FLA hook path

Eight issues reproduced by parallel reviewers against 6ce495a; all
fixed and covered by regression tests. 45 pytest cases pass (was 36);
end-to-end Qwen3.5_MoE modeling-import drill still loads all five
fast-path symbols.

P1 fixes:

1. TileLang loses the Qwen-family guard on the normal FLA hook path
   (10/12 reviewers, reproduced with allenai/OLMo-Hybrid-1B). The
   hook unconditionally installed tilelang for any FLA-using model.
   - Threaded `model_name` through `_install_fast_path_hooks(event_queue,
     model_name)`.
   - `_fla_install` now gates tilelang on
     `_model_wants_tilelang(model_name)` AND a successful FLA install.

2. TileLang repair `--force-reinstall` (without `--no-deps`) could
   replace `torch==2.12.0+cu130` with `torch==2.12.0`. Split repair
   into TWO steps:
     step 1: `--force-reinstall --no-deps apache-tvm-ffi==0.1.9`
     step 2: regular install of tilelang + apache-tvm-ffi
   Step 1 surgically downgrades the broken package; step 2 resolves
   missing transitive deps (z3-solver, ml-dtypes) without
   --force-reinstall, so it never replaces torch.

3. Hook could return True after the installer's deep import probe
   failed: when pip exits 0 but `import fla.modules` raises, the old
   wrapper re-called `original()` (transformers' metadata check) and
   trusted it. Refactored:
     - `_ensure_flash_linear_attention_unconditional(...) -> bool`
     - `_ensure_tilelang_backend_unconditional(...) -> bool`
   The wrapper now uses the installer's bool directly.

4. SSM models (Nemotron-H, Falcon-H1, Granite-H) use
   `lazy_load_kernel("causal-conv1d")` and never call
   `is_causal_conv1d_available()`, so the hook never fires for them.
   The orchestrator now always runs `_ensure_causal_conv1d_fast_path`
   outside the hook-mode if/else.

P2 fixes:

5. `_rebind_in_already_imported_modules` invoked transformers' lazy
   module `__getattr__` (hundreds of "Accessing X from .models..."
   warnings, ~3.4s overhead). Switched to `module.__dict__.get(...)`
   which only sees real module-level bindings.

6. TileLang installed even when FLA was skipped (Torch <2.7) or
   failed (timeout, post-install probe failed). Now gated on the
   installer's bool return.

7. TileLang repair was skipped when FLA was already True but tilelang
   missing or apache-tvm-ffi on the broken list. Added an optional
   `post_available_fn` to the wrapper; the FLA hook's
   `_fla_post_available` runs `_ensure_tilelang_backend_unconditional`
   when (model wants tilelang) AND (tilelang missing OR tvm-ffi broken).

8. `_flash_linear_attention_importable()` only checks deep import,
   not version. Added `_flash_linear_attention_current()` that
   compares against the pinned `flash-linear-attention==0.5.0` /
   `fla-core==0.5.0`; older versions trigger `--force-reinstall
   --no-deps` so torch stays untouched.

Helpers extracted to keep the surface tight:
  - `_pip_install_cmd(*args)` builds `uv pip install` or
    `python -m pip install` depending on uv availability.
  - `_run_pip(cmd, event_queue, label)` runs a pip command with
    timeout / failure handling and a status emission.

Regression tests added:

  - test_hook_does_not_install_tilelang_for_non_qwen_fla_model
  - test_hook_does_install_tilelang_for_qwen35
  - test_tilelang_repair_does_not_touch_torch_cuda_stack
  - test_hook_trusts_installer_bool_not_metadata
  - test_rebind_does_not_trigger_module_getattr
  - test_hook_skips_tilelang_when_fla_install_is_skipped
  - test_hook_runs_tilelang_repair_when_fla_already_true
  - test_fla_installer_force_reinstalls_when_older_version_present
  - test_run_training_process_eagerly_installs_causal_conv1d_in_normal_mode

Existing tests updated for the new `_install_fast_path_hooks` signature
and the two-step tilelang repair flow.

End-to-end re-verified against transformers.models.qwen3_5_moe:
PRE_STATE fla=False, hook fires for both gates, FLA + tilelang +
causal-conv1d install, all 5 fast-path symbols non-None.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fix double-install of tilelang on the FLA hook install path

Backend CI surfaced a test-isolation bug introduced by the
post_available_fn mechanism for finding #7. The wrapper ran
`post_available_fn` in BOTH paths (install ran AND gate already True),
but `_fla_install` already chains tilelang on the install path, so the
post-available step then called tilelang install AGAIN.

This was masked locally because tilelang was installed in the
workspace venv (post_available short-circuited on
`_tilelang_importable()` returning True). CI starts with no tilelang,
so the second call actually fired and the mock recorded two calls.

Fix: only run `post_available_fn` when the install path did NOT run.
That preserves the finding #7 semantics (tilelang repair when FLA
already True but tilelang missing or tvm-ffi broken) without
duplicating the chained install on the gate-was-False path.

Also tightened `test_hook_skips_install_when_gate_already_true` to
monkeypatch `_tilelang_importable=True` and
`_installed_tvm_ffi_version=0.1.9` so it stays a pure "no install at
all" test regardless of the venv's actual state.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* ci: retrigger Mac Studio GGUF after transient HF DNS resolve flake

* studio: skip tilelang on HIP / ROCm torch (Strix Halo crash report)

h34v3nzc0dex tested PR 5434 on Strix Halo (gfx1151, ROCm 7.13,
torch 2.11.0+rocm7.13.0) and hit a hard regression:

  File ".../fla/ops/common/backends/tilelang/__init__.py", line 92,
    in chunk_bwd_dqkwg
  File ".../tilelang/jit/kernel.py", line 137, in __init__
  File ".../tilelang/tileop/gemm/__init__.py", line 143,
    in _select_gemm_instruction
  tvm.error.InternalError: Check failed: (0) is false:
    Unsupported target for gemm:
    hip -keys=hip,gpu -mcpu=gfx1151 ...

`tilelang==0.1.8` ships no HIP GEMM instruction; `_select_gemm_instruction`
raises at lower-time, not import-time. So:
  - pip install succeeds
  - `import tilelang` succeeds
  - `TileLangBackend.is_available()` returns True
  - FLA's dispatcher picks TileLang for `chunk_bwd_dqkwg`
  - training subprocess dies at first GDN backward, no graceful fallback

The PR's existing platform gate (`_tilelang_platform_supported`)
checked only `sys.platform == "linux"` and `platform.machine()`, both
of which look identical on a ROCm box.

Fix has two layers:

1. INSTALL GATE: new `_torch_has_hip()` helper checks
   `torch.version.hip is not None`. `_tilelang_platform_supported`
   now returns False on HIP torch, so the install never fires.

2. RUNTIME GATE: even with the install skipped, a user could have
   tilelang already present (e.g. venv carried over from a CUDA box).
   `_install_fast_path_hooks` now calls
   `os.environ.setdefault("FLA_TILELANG", "0")` when HIP is detected,
   which is the env-var FLA's `TileLangBackend` already honors. Users
   who know they have a HIP-aware tilelang fork can override by
   setting `FLA_TILELANG=1` explicitly.

This costs nothing on CUDA (the gate is a no-op when
`torch.version.hip is None`), and removes the crash for AMD users.
The benchmark numbers in the PR description (1.43x on B200 sm_100)
are not affected.

The other halves of the PR are confirmed working on gfx1151 by the
same report:
  - `flash-linear-attention 0.5.0` runs at production scale
    (B=1 T=8192 H=16 K=128 V=128 and others) with no patches.
  - `causal-conv1d` runs at the shapes the fast-path gate cares
    about. (A separate Ubuntu 24.04 `--gcc-install-dir` build
    workaround is needed for the source-build path; that mirrors
    bbf004c's llama.cpp fix and is out of scope here.)

Tests added:
  - test_tilelang_platform_unsupported_on_hip_torch
  - test_tilelang_install_skipped_on_hip_torch
  - test_install_fast_path_hooks_sets_fla_tilelang_zero_on_hip
  - test_install_fast_path_hooks_respects_user_fla_tilelang_override
  - test_install_fast_path_hooks_does_not_set_fla_tilelang_on_cuda

Total 50 passing (was 45).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* ci: retrigger Windows Studio UI after transient Playwright tab-lookup flake

* studio: auto-discover FLA-using model types from installed transformers

Drop the hand-maintained `_TILELANG_MODEL_SUBSTRINGS` tuple
(qwen3.5 / qwen3_5 / qwen3.6 / qwen3_6 / qwen3-next / qwen3_next)
and derive the allowlist by scanning the installed
`transformers/models/*/modeling_*.py` for `from fla.` imports.

A model "wants tilelang" iff its modeling file imports an FLA op,
which is the same signal `is_flash_linear_attention_available()` is
the runtime test for. The scan happens once per worker subprocess
and is cached for the process lifetime; an empty result (eg
transformers not importable) means "no tilelang pre-install" --
the FLA runtime hook still drives the install via the gate when
the loaded model actually probes it.

Verified against the live installed transformers, the auto-derived
set is {qwen3_5, qwen3_5_moe, qwen3_next}, with `_model_wants_tilelang`
matching the HF Hub names `unsloth/Qwen3.5-2B`, `Qwen/Qwen3.5-MoE-A3B`,
`mlx-community/qwen3-next-80b`, and correctly rejecting Llama,
Mistral, Nemotron-H, Falcon-H1, etc. Future GDN models (Qwen3.7,
OLMo-Hybrid-FA, ...) are picked up automatically once they ship in
transformers; no further worker edits needed.

Also trim docstrings / comments through the FLA / tilelang / HIP /
hook block: constants get 1-line trailing comments, function
docstrings collapse to 1-3 lines, and the fast-path-hooks banner
shrinks from a 27-line block to 4 lines. The file drops from 2847
to 2630 lines without losing the load-bearing WHY notes
(--no-deps protects torch; `__dict__.get` avoids lazy-module
__getattr__; two-step tvm-ffi repair keeps torch off the dep
graph; HIP setdefault disables FLA's TileLang dispatch even with
tilelang already installed).

7 new tests (50 -> 57 total): discovery returns only FLA-using
model_types; discovery cache reuse; missing transformers handled;
OSError on a modeling file is non-fatal; `_model_wants_tilelang`
matches real HF repo names across separator variants; empty
discovery -> always False; normalization across `-`, `.`, `/`,
space.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* test: hermetize the non-allowlist hook test against transformers 5.4.0+

transformers 5.4.0 added `olmo_hybrid` as an FLA-using model_type, so
the auto-discovered allowlist now includes it -- and the test's prior
choice of `allenai/OLMo-Hybrid-1B` as a "non-Qwen FLA-only" example
became an allowlist member. CI on Python 3.11 / 3.13 caught this.

Swap to a guaranteed-not-in-allowlist fake model_name AND patch
_discover_fla_model_types to a known {qwen3_5, qwen3_5_moe, qwen3_next}
set so the test stays valid as upstream transformers adds new
FLA-using architectures.

Renames the test to reflect the actual semantic under test:
"outside-allowlist -> no tilelang".

* ci: retrigger Windows Studio API after llama.cpp prebuilt staging WinError 5 flake

* tests: move MLX smoke gate changes to dedicated PR #5537

The seven MLX smoke commits in this PR's history (_on_step grad_norm,
max_grad_value pin, loss + round-trip gates) are unrelated to the
FLA / tilelang work. They now live in #5537 so this PR's diff is
limited to the studio worker installer changes.

Net effect on tests/studio/run_real_mlx_smoke.py vs main: zero.

* studio: friendlier install banners (drop hook / gate-name jargon)

User-visible status text now reads:
  Installing flash-linear-attention==<ver> for faster training...
  Installing TileLang==<ver> for faster training...
  Installing causal-conv1d for faster training...
  Installing flash-attn for faster training...

Removed the transient "Hook fired for is_flash_linear_attention_available;
installing kernel..." banner — the install banner that immediately follows
already tells the user what is happening, in plain English.

The internal logger.info messages (server-side log) still carry the
gate names + "Hook fired ..." for debugging.

* [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>
2026-05-18 03:49:06 -07:00

2729 lines
101 KiB
Python

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""
Training subprocess entry point.
Each training job runs in a fresh subprocess (mp.get_context("spawn")).
This gives us a clean Python interpreter with no stale module state —
solving the transformers version-switching problem completely.
Pattern follows core/data_recipe/jobs/worker.py.
"""
from __future__ import annotations
import structlog
from loggers import get_logger
import math
import os
import shutil
import sys
import time
import traceback
import subprocess as _sp
from pathlib import Path
from typing import Any, Callable
logger = get_logger(__name__)
from utils.hardware import apply_gpu_ids
from utils.wheel_utils import (
direct_wheel_url,
flash_attn_wheel_url,
has_blackwell_gpu,
install_wheel,
probe_torch_wheel_env,
url_exists,
)
def _output_dir_from_resume_checkpoint(
resume_from_checkpoint: str | None,
) -> str | None:
if not resume_from_checkpoint:
return None
path = Path(resume_from_checkpoint)
return str(path.parent if path.name.startswith("checkpoint-") else path)
_CAUSAL_CONV1D_RELEASE_TAG = "v1.6.1.post4"
_CAUSAL_CONV1D_PACKAGE_VERSION = "1.6.1"
_MAMBA_SSM_RELEASE_TAG = "v2.3.1"
_MAMBA_SSM_PACKAGE_VERSION = "2.3.1"
_FLASH_ATTN_RUNTIME_MIN_SEQ_LEN = 32768
_FLASH_ATTN_SKIP_ENV = "UNSLOTH_STUDIO_SKIP_FLASHATTN_INSTALL"
# apache-tvm-ffi 0.1.10/0.1.11 crash Triton with "CUDA: misaligned address" on sm_100.
_TILELANG_PACKAGE_VERSION = "0.1.8"
_APACHE_TVM_FFI_PACKAGE_VERSION = "0.1.9"
_TILELANG_SKIP_ENV = "UNSLOTH_STUDIO_SKIP_TILELANG_INSTALL"
# Pin both so plain pip cannot silently upgrade torch under the worker (fla-core needs torch>=2.7).
_FLA_PACKAGE_VERSION = "0.5.0"
_FLA_CORE_PACKAGE_VERSION = "0.5.0"
_FLA_SKIP_ENV = "UNSLOTH_STUDIO_SKIP_FLA_INSTALL"
# `--no-deps` saves torch but loses fla-core's transitive deps; `packaging` is also undeclared upstream.
_FLA_RUNTIME_DEPS = ("einops", "packaging", "triton")
_FLA_MIN_TORCH = (2, 7)
_FLA_MIN_PYTHON = (3, 10)
# tilelang 0.1.8 ships wheels only for these Linux arches and macOS arm64; never fall back to its 93MB sdist.
_TILELANG_SUPPORTED_LINUX_MACHINES = frozenset(("x86_64", "amd64", "aarch64", "arm64"))
_TILELANG_INSTALL_TIMEOUT_S = 600
_TVM_FFI_BROKEN_VERSIONS = ("0.1.10", "0.1.11")
_FAST_PATH_HOOKS_SKIP_ENV = "UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS"
def _model_wants_causal_conv1d(model_name: str) -> bool:
name = model_name.lower()
return any(
key in name
for key in (
"qwen3.5",
"qwen3_5",
"qwen3.6",
"qwen3_6",
"qwen3-next",
"qwen3_next",
"nemotron_h",
"nemotron-h",
"nemotron-3-nano",
"falcon_h1",
"falcon-h1",
"granite-4.0-h",
"granitemoehybrid",
"lfm2",
)
)
def _hipcc_gcc_install_dir() -> str | None:
"""Return the highest-numbered ``/usr/lib/gcc/x86_64-linux-gnu/<N>`` that has
BOTH the gcc runtime dir AND the corresponding ``/usr/include/c++/<N>`` C++
headers, or ``None`` if no match (or non-Linux / non-x86_64).
Ubuntu 24.04 ships ``/usr/lib/gcc/x86_64-linux-gnu/14/`` (gcc-14 runtime
objects) but does NOT ship ``/usr/include/c++/14`` in its default apt set;
libstdc++ headers come from ``libstdc++-13-dev``. ROCm clang-20 picks the
highest-numbered runtime dir by default, finds no ``<cstdlib>``, and the
HIP source build fails with::
/opt/rocm-X.Y/lib/llvm/lib/clang/20/include/__clang_hip_runtime_wrapper.h:112:10:
fatal error: 'cstdlib' file not found
Returning a path lets the caller pass ``--gcc-install-dir=<path>`` to clang
via ``HIPCC_COMPILE_FLAGS_APPEND``. Mirrors the same loop ``bbf004c`` added
to ``studio/setup.sh`` for the llama.cpp HIP build branch (PR #5301).
"""
if not sys.platform.startswith("linux"):
return None
import platform as _platform
if _platform.machine().lower() != "x86_64":
return None
for _ver in (14, 13, 12, 11):
_runtime = f"/usr/lib/gcc/x86_64-linux-gnu/{_ver}/include"
_headers = f"/usr/include/c++/{_ver}"
if os.path.isdir(_runtime) and os.path.isdir(_headers):
return f"/usr/lib/gcc/x86_64-linux-gnu/{_ver}"
return None
def _install_package_wheel_first(
*,
event_queue: Any,
import_name: str,
display_name: str,
pypi_name: str,
pypi_version: str | None = None,
filename_prefix: str | None = None,
release_tag: str | None = None,
release_base_url: str | None = None,
wheel_url_builder: Callable[[dict[str, str] | None], str | None] | None = None,
pypi_spec: str | None = None,
pypi_status_message: str | None = None,
) -> bool:
try:
__import__(import_name)
logger.info("%s already installed", display_name)
return True
except ImportError:
pass
env = probe_torch_wheel_env(timeout = 30)
if wheel_url_builder is not None:
wheel_url = wheel_url_builder(env)
else:
wheel_url = direct_wheel_url(
filename_prefix = filename_prefix,
package_version = pypi_version,
release_tag = release_tag,
release_base_url = release_base_url,
env = env,
)
if wheel_url is None:
logger.info("No compatible %s wheel candidate", display_name)
elif url_exists(wheel_url):
_send_status(event_queue, f"Installing {display_name} for faster training...")
for installer, result in install_wheel(
wheel_url,
python_executable = sys.executable,
use_uv = bool(shutil.which("uv")),
run = _sp.run,
):
if result.returncode == 0:
logger.info("Installed prebuilt %s wheel successfully", display_name)
return True
logger.warning(
"%s failed to install %s wheel:\n%s",
installer,
display_name,
result.stdout,
)
else:
logger.info("No published %s wheel found: %s", display_name, wheel_url)
is_hip = env and env.get("hip_version")
if is_hip and not shutil.which("hipcc"):
logger.error(
"%s requires hipcc for source compilation on ROCm. "
"Install the ROCm HIP SDK: https://rocm.docs.amd.com",
display_name,
)
_send_status(
event_queue,
f"{display_name}: hipcc not found (ROCm HIP SDK required)",
)
return False
if pypi_spec is None:
pypi_spec = f"{pypi_name}=={pypi_version}"
if pypi_status_message is None:
if is_hip:
pypi_status_message = (
f"Compiling {display_name} from source for ROCm "
"(this may take several minutes)..."
)
else:
pypi_status_message = (
f"Installing {display_name} from PyPI for faster training..."
)
_send_status(event_queue, pypi_status_message)
# Prefer uv for faster dependency resolution when available
plain_pypi_install = pypi_version is None
if plain_pypi_install:
if shutil.which("uv"):
pypi_cmd = [
"uv",
"pip",
"install",
"--python",
sys.executable,
pypi_spec,
]
else:
pypi_cmd = [sys.executable, "-m", "pip", "install", pypi_spec]
else:
if shutil.which("uv"):
pypi_cmd = [
"uv",
"pip",
"install",
"--python",
sys.executable,
"--no-build-isolation",
"--no-deps",
]
# Avoid stale cache artifacts from partial HIP source builds
if is_hip:
pypi_cmd.append("--no-cache")
pypi_cmd.append(pypi_spec)
else:
pypi_cmd = [
sys.executable,
"-m",
"pip",
"install",
"--no-build-isolation",
"--no-deps",
"--no-cache-dir",
pypi_spec,
]
# Source compilation on ROCm can take 10-30 minutes; use a generous
# timeout. Non-HIP installs preserve the pre-existing "no timeout"
# behaviour so unrelated slow installs (e.g. causal-conv1d source
# build on Linux aarch64 or unsupported torch/CUDA combinations)
# are not aborted at 5 minutes by this PR.
_run_kwargs: dict[str, Any] = {
"stdout": _sp.PIPE,
"stderr": _sp.STDOUT,
"text": True,
}
if is_hip:
_run_kwargs["timeout"] = 1800
# On Ubuntu 24.04 + ROCm clang-20, the HIP source build (causal-conv1d,
# mamba-ssm source fallback, flash-attn source fallback) defaults to
# /usr/lib/gcc/x86_64-linux-gnu/14/ which has the runtime dir but no
# /usr/include/c++/14 headers, and dies at:
# __clang_hip_runtime_wrapper.h:112:10:
# fatal error: 'cstdlib' file not found
# Inject --gcc-install-dir for a gcc whose C++ headers actually exist.
# Respect any pre-existing --gcc-install-dir in HIPCC_COMPILE_FLAGS_APPEND
# (user knows best); otherwise append. Mirrors the same fix bbf004c
# added to studio/setup.sh for the llama.cpp HIP build (PR #5301).
_existing_flags = os.environ.get("HIPCC_COMPILE_FLAGS_APPEND", "")
if "--gcc-install-dir" not in _existing_flags:
_gcc_dir = _hipcc_gcc_install_dir()
if _gcc_dir is not None:
_appended = (f"{_existing_flags} --gcc-install-dir={_gcc_dir}").strip()
_env = _run_kwargs.get("env", os.environ).copy()
_env["HIPCC_COMPILE_FLAGS_APPEND"] = _appended
_run_kwargs["env"] = _env
logger.info(
"HIP source build for %s: appended "
"--gcc-install-dir=%s to HIPCC_COMPILE_FLAGS_APPEND",
display_name,
_gcc_dir,
)
try:
result = _sp.run(pypi_cmd, **_run_kwargs)
except _sp.TimeoutExpired:
logger.error(
"%s installation timed out after %ds",
display_name,
_run_kwargs.get("timeout"),
)
_send_status(
event_queue,
f"{display_name} installation timed out after "
f"{_run_kwargs.get('timeout')}s",
)
return False
if result.returncode != 0:
if is_hip:
# Surface a clear error for ROCm source build failures
error_lines = (result.stdout or "").strip().splitlines()
snippet = "\n".join(error_lines[-5:]) if error_lines else "(no output)"
logger.error(
"Failed to compile %s for ROCm:\n%s",
display_name,
result.stdout,
)
_send_status(
event_queue,
f"Failed to compile {display_name} for ROCm. "
"Check that hipcc and ROCm development headers are installed.\n"
f"{snippet}",
)
else:
logger.error(
"Failed to install %s from PyPI:\n%s",
display_name,
result.stdout,
)
return False
if is_hip:
logger.info("Compiled and installed %s from source for ROCm", display_name)
else:
logger.info("Installed %s from PyPI", display_name)
return True
def _ensure_causal_conv1d_fast_path(event_queue: Any, model_name: str) -> None:
if not _model_wants_causal_conv1d(model_name):
return
_install_package_wheel_first(
event_queue = event_queue,
import_name = "causal_conv1d",
display_name = "causal-conv1d",
pypi_name = "causal-conv1d",
pypi_version = _CAUSAL_CONV1D_PACKAGE_VERSION,
filename_prefix = "causal_conv1d",
release_tag = _CAUSAL_CONV1D_RELEASE_TAG,
release_base_url = "https://github.com/Dao-AILab/causal-conv1d/releases/download",
)
def _installed_torch_version_tuple() -> tuple[int, int] | None:
"""Return ``(major, minor)`` of the installed torch, else None."""
try:
from importlib.metadata import version as _pkg_version
raw = _pkg_version("torch").split("+", 1)[0]
parts = raw.split(".")
return (int(parts[0]), int(parts[1]))
except Exception:
return None
def _flash_linear_attention_importable() -> bool:
"""Catch any exception (not just ImportError) so a broken native lib doesn't abort the worker."""
try:
import fla.modules # noqa: F401
import fla.ops.gated_delta_rule # noqa: F401
return True
except Exception as exc:
logger.warning(
"flash-linear-attention is not importable; continuing with install/fallback: %s",
exc,
)
return False
def _flash_linear_attention_current(already_importable: bool | None = None) -> bool:
"""True iff FLA imports AND is at the pinned version (older FLA lacks gated_delta_rule kernels)."""
if already_importable is None:
already_importable = _flash_linear_attention_importable()
if not already_importable:
return False
try:
from importlib.metadata import version as _pkg_version
from packaging.version import Version
fla_v = Version(_pkg_version("flash-linear-attention"))
core_v = Version(_pkg_version("fla-core"))
return fla_v >= Version(_FLA_PACKAGE_VERSION) and core_v >= Version(
_FLA_CORE_PACKAGE_VERSION
)
except Exception as exc:
logger.warning(
"flash-linear-attention importable but version check failed; treating as stale: %s",
exc,
)
return False
def _ensure_flash_linear_attention_unconditional(event_queue: Any) -> bool:
"""Install pinned FLA + fla-core with --no-deps. Returns True iff importable post-call."""
if os.getenv(_FLA_SKIP_ENV) == "1":
return False
if sys.version_info < _FLA_MIN_PYTHON:
logger.info(
"Skipping flash-linear-attention install: requires Python >= %d.%d, have %s",
_FLA_MIN_PYTHON[0],
_FLA_MIN_PYTHON[1],
sys.version.split()[0],
)
return False
torch_ver = _installed_torch_version_tuple()
if torch_ver is not None and torch_ver < _FLA_MIN_TORCH:
_send_status(
event_queue,
(
f"Skipping flash-linear-attention install: fla-core requires "
f"torch>={_FLA_MIN_TORCH[0]}.{_FLA_MIN_TORCH[1]}, have "
f"{torch_ver[0]}.{torch_ver[1]}"
),
)
return False
# Probe once; reuse result so the --force-reinstall decision and the short-circuit
# share the same call count (stable for tests).
already_importable = _flash_linear_attention_importable()
if already_importable and _flash_linear_attention_current(already_importable = True):
logger.info("flash-linear-attention already importable at the pinned version")
return True
_send_status(
event_queue,
f"Installing flash-linear-attention=={_FLA_PACKAGE_VERSION} for faster training...",
)
# `--no-deps` blocks the silent torch upgrade; we bring the non-torch runtime deps in by hand.
specs = [
*_FLA_RUNTIME_DEPS,
f"fla-core=={_FLA_CORE_PACKAGE_VERSION}",
f"flash-linear-attention=={_FLA_PACKAGE_VERSION}",
]
extra_args = ["--no-deps"]
if already_importable:
# Older FLA already imported; pip skips reinstall without this flag.
extra_args.append("--force-reinstall")
if shutil.which("uv"):
pypi_cmd = [
"uv",
"pip",
"install",
"--python",
sys.executable,
*extra_args,
*specs,
]
else:
pypi_cmd = [
sys.executable,
"-m",
"pip",
"install",
*extra_args,
*specs,
]
try:
result = _sp.run(
pypi_cmd,
stdout = _sp.PIPE,
stderr = _sp.STDOUT,
text = True,
timeout = _TILELANG_INSTALL_TIMEOUT_S,
)
except _sp.TimeoutExpired:
logger.warning("flash-linear-attention install timed out; continuing")
_send_status(
event_queue, "flash-linear-attention install timed out; continuing"
)
return False
if result.returncode != 0:
logger.warning(
"flash-linear-attention install failed (continuing on torch fallback):\n%s",
result.stdout,
)
_send_status(
event_queue,
"flash-linear-attention install failed; continuing without it",
)
return False
# pip can exit 0 with a missing transitive runtime dep; verify the import.
if not _flash_linear_attention_importable():
_send_status(
event_queue,
"flash-linear-attention installed but is not importable; continuing without it",
)
return False
logger.info("Installed flash-linear-attention for the FLA fast path")
return True
def _ensure_flash_linear_attention(event_queue: Any, model_name: str) -> None:
"""Legacy model-name-gated FLA install, used when UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS=1."""
if not _model_wants_tilelang(model_name):
return
_ensure_flash_linear_attention_unconditional(event_queue)
_SSM_MODEL_SUBSTRINGS = (
"nemotron_h",
"nemotron-h",
"nemotron-3-nano",
"falcon_h1",
"falcon-h1",
"granite-4.0-h",
"granitemoehybrid",
)
def _ensure_mamba_ssm(event_queue: Any, model_name: str) -> None:
if not any(sub in model_name.lower() for sub in _SSM_MODEL_SUBSTRINGS):
return
logger.info("SSM model detected; setting up mamba-ssm after causal-conv1d")
_install_package_wheel_first(
event_queue = event_queue,
import_name = "mamba_ssm",
display_name = "mamba-ssm",
pypi_name = "mamba-ssm",
pypi_version = _MAMBA_SSM_PACKAGE_VERSION,
filename_prefix = "mamba_ssm",
release_tag = _MAMBA_SSM_RELEASE_TAG,
release_base_url = "https://github.com/state-spaces/mamba/releases/download",
)
# Auto-derived from installed transformers: model_types whose modeling_*.py imports `from fla.*`.
# Cached per process. Empty when transformers can't be inspected -> we skip tilelang pre-install
# (the FLA Triton path still runs via the runtime hook).
_TRANSFORMERS_FLA_MODEL_TYPES_CACHE: frozenset[str] | None = None
_MODEL_NAME_SEP_CHARS = ("-", ".", "/", " ")
def _discover_fla_model_types() -> frozenset[str]:
"""Model_types in the installed transformers whose modeling file imports `from fla.*`."""
global _TRANSFORMERS_FLA_MODEL_TYPES_CACHE
if _TRANSFORMERS_FLA_MODEL_TYPES_CACHE is not None:
return _TRANSFORMERS_FLA_MODEL_TYPES_CACHE
found: set[str] = set()
try:
import transformers
models_root = Path(transformers.__file__).parent / "models"
for modeling in models_root.glob("*/modeling_*.py"):
try:
src = modeling.read_text(encoding = "utf-8", errors = "ignore")
except OSError:
continue
if "from fla." in src:
found.add(modeling.parent.name)
except Exception as exc:
logger.debug("FLA model-type discovery skipped: %s", exc)
_TRANSFORMERS_FLA_MODEL_TYPES_CACHE = frozenset(found)
return _TRANSFORMERS_FLA_MODEL_TYPES_CACHE
def _model_wants_tilelang(model_name: str) -> bool:
"""True iff model_name normalizes to contain a discovered FLA model_type."""
types = _discover_fla_model_types()
if not types:
return False
name = model_name.lower()
for sep in _MODEL_NAME_SEP_CHARS:
name = name.replace(sep, "_")
return any(t in name for t in types)
def _installed_tvm_ffi_version() -> str | None:
"""Installed apache-tvm-ffi version, or None if missing/unimportable."""
try:
from importlib.metadata import version as _pkg_version
return _pkg_version("apache-tvm-ffi")
except Exception:
return None
def _tilelang_importable() -> bool:
"""Catch any exception (not just ImportError) so a broken native lib doesn't abort the worker."""
try:
import tilelang # noqa: F401
import tvm_ffi # noqa: F401
return True
except Exception as exc:
logger.warning(
"tilelang/tvm_ffi is not importable; continuing with install/fallback: %s",
exc,
)
return False
def _torch_has_hip() -> bool:
"""True iff torch is a ROCm build; `torch.version.hip` is the only reliable signal on x86_64 ROCm."""
try:
import torch as _torch
return getattr(_torch.version, "hip", None) is not None
except Exception:
return False
def _tilelang_platform_supported() -> bool:
"""True iff a tilelang 0.1.8 wheel will load: Linux x86_64/aarch64, non-HIP torch.
HIP excluded because tilelang 0.1.8 has no HIP GEMM instruction and crashes mid-backward.
"""
import platform as _platform
if not sys.platform.startswith("linux"):
return False
if _platform.machine().lower() not in _TILELANG_SUPPORTED_LINUX_MACHINES:
return False
if _torch_has_hip():
return False
return True
def _pip_install_cmd(*args: str) -> list[str]:
"""`uv pip install` if uv is on PATH, else `python -m pip install`."""
if shutil.which("uv"):
return ["uv", "pip", "install", "--python", sys.executable, *args]
return [sys.executable, "-m", "pip", "install", *args]
def _run_pip(cmd: list[str], event_queue: Any, label: str) -> bool:
"""Run a pip install and surface success/failure via status events."""
try:
result = _sp.run(
cmd,
stdout = _sp.PIPE,
stderr = _sp.STDOUT,
text = True,
timeout = _TILELANG_INSTALL_TIMEOUT_S,
)
except _sp.TimeoutExpired:
logger.warning("%s install timed out; continuing", label)
_send_status(event_queue, f"{label} install timed out; continuing")
return False
if result.returncode != 0:
logger.warning(
"%s install failed (continuing without it):\n%s", label, result.stdout
)
_send_status(event_queue, f"{label} install failed; continuing")
return False
return True
def _ensure_tilelang_backend_unconditional(event_queue: Any) -> bool:
"""Install pinned tilelang + apache-tvm-ffi; two-step repair if a broken tvm-ffi is present.
Returns True iff both import post-call. Step 1 surgically downgrades a broken tvm-ffi
with --force-reinstall --no-deps so torch / CUDA stay untouched; step 2 is a regular
install for missing transitive deps. Bypass via UNSLOTH_STUDIO_SKIP_TILELANG_INSTALL=1.
"""
if os.getenv(_TILELANG_SKIP_ENV) == "1":
return False
if sys.version_info < _FLA_MIN_PYTHON:
logger.info(
"Skipping tilelang install: requires Python >= %d.%d, have %s",
_FLA_MIN_PYTHON[0],
_FLA_MIN_PYTHON[1],
sys.version.split()[0],
)
return False
if not _tilelang_platform_supported():
import platform as _platform
logger.info(
"Skipping tilelang install: no prebuilt wheel for %s/%s",
sys.platform,
_platform.machine(),
)
return False
existing_tvm_ffi = _installed_tvm_ffi_version()
needs_repair = existing_tvm_ffi in _TVM_FFI_BROKEN_VERSIONS
if not needs_repair and _tilelang_importable():
logger.info("tilelang + apache-tvm-ffi already installed")
return True
# Step 1: --no-deps keeps --force-reinstall from touching torch/CUDA via the dep graph.
if needs_repair:
logger.info(
"Forcing apache-tvm-ffi downgrade: %s is on the broken list",
existing_tvm_ffi,
)
_send_status(
event_queue,
(
f"Downgrading apache-tvm-ffi {existing_tvm_ffi} -> "
f"{_APACHE_TVM_FFI_PACKAGE_VERSION} (broken-versions list)"
),
)
repair_cmd = _pip_install_cmd(
"--only-binary=:all:",
"--force-reinstall",
"--no-deps",
f"apache-tvm-ffi=={_APACHE_TVM_FFI_PACKAGE_VERSION}",
)
if not _run_pip(repair_cmd, event_queue, "TileLang backend repair"):
return False
# Step 2: regular install pulls in transitive deps (z3-solver, ml-dtypes) without touching torch.
_send_status(
event_queue,
f"Installing TileLang=={_TILELANG_PACKAGE_VERSION} for faster training...",
)
install_cmd = _pip_install_cmd(
"--only-binary=:all:",
f"apache-tvm-ffi=={_APACHE_TVM_FFI_PACKAGE_VERSION}",
f"tilelang=={_TILELANG_PACKAGE_VERSION}",
)
if not _run_pip(install_cmd, event_queue, "TileLang backend"):
return False
# pip can exit 0 while a native lib (libz3.so) is missing; verify the import.
if not _tilelang_importable():
_send_status(
event_queue,
"TileLang backend installed but is not importable; continuing on the FLA Triton path",
)
return False
logger.info("Installed TileLang backend for FLA fast path")
return True
def _ensure_tilelang_backend(event_queue: Any, model_name: str) -> None:
"""Legacy substring-gated tilelang installer (opt-out path)."""
if not _model_wants_tilelang(model_name):
return
_ensure_tilelang_backend_unconditional(event_queue)
# ── Fast-path hooks ──
# Wrap transformers' is_{flash_linear_attention,causal_conv1d}_available so the first call
# (at modeling import time) drives the install. Any model that queries the gate gets the
# install; models that never query it (Llama, Gemma, dense Qwen) pay nothing.
# UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS=1 falls back to the legacy substring path.
def _rebind_in_already_imported_modules(
*, attr_name: str, old_obj: Any, new_obj: Any
) -> int:
"""Rebind `attr_name -> new_obj` in every module that already imported `old_obj`.
`from X import Y` creates a local binding that reassigning X.Y won't reach.
Uses `__dict__.get` (not `getattr`) to skip lazy `__getattr__` aliases.
"""
count = 0
missing = object()
for mod_name, mod in list(sys.modules.items()):
if mod is None:
continue
module_dict = getattr(mod, "__dict__", None)
if not isinstance(module_dict, dict):
continue
existing = module_dict.get(attr_name, missing)
if existing is old_obj:
try:
setattr(mod, attr_name, new_obj)
count += 1
except Exception as exc:
logger.debug("Could not rebind %s in %s: %s", attr_name, mod_name, exc)
return count
def _install_fast_path_hooks(event_queue: Any, model_name: str) -> None:
"""Hook transformers' is_*_available gates so the first call drives the install.
Idempotent. UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS=1 falls back to the substring gate.
"""
if os.getenv(_FAST_PATH_HOOKS_SKIP_ENV) == "1":
logger.info("Fast-path hooks disabled via env; using substring fallback")
return
# On HIP torch, even already-installed tilelang crashes FLA's TileLang dispatch.
# User can override with FLA_TILELANG=1.
if _torch_has_hip() and os.environ.get("FLA_TILELANG") is None:
os.environ["FLA_TILELANG"] = "0"
logger.info(
"HIP/ROCm torch detected; setting FLA_TILELANG=0 (no HIP GEMM in tilelang 0.1.8)"
)
try:
from transformers.utils import import_utils as _iu
except Exception as exc:
logger.warning(
"transformers.utils.import_utils not importable; skipping fast-path hooks: %s",
exc,
)
return
def _make_wrapper(
original: Callable[[], bool],
install_fn: Callable[[Any], bool],
gate_name: str,
post_available_fn: Callable[[Any], None] | None = None,
) -> Callable[[], bool]:
state = {"installed": False}
def wrapper() -> bool:
if state["installed"]:
return original()
try:
original.cache_clear() # defensive; worker subprocess is fresh
except AttributeError:
pass
ok = original()
ran_install = False
if not ok:
ran_install = True
logger.info("Hook fired for %s; triggering install", gate_name)
try:
ok = bool(install_fn(event_queue))
except Exception as exc:
logger.warning(
"%s install raised: %s; falling back to torch", gate_name, exc
)
ok = False
logger.info("%s hook done; available=%s", gate_name, ok)
# post_available_fn handles "gate already True but ancillary kernel broken" (e.g. tilelang
# missing while FLA imports fine); skip when install_fn already chained the follow-up.
if ok and not ran_install and post_available_fn is not None:
try:
post_available_fn(event_queue)
except Exception as exc:
logger.warning(
"%s post-available step raised: %s; continuing", gate_name, exc
)
state["installed"] = True
return ok
wrapper.__wrapped__ = original # type: ignore[attr-defined]
wrapper.cache_clear = getattr(original, "cache_clear", lambda: None) # type: ignore[attr-defined]
return wrapper
def _fla_install(eq: Any) -> bool:
# FLA alone ~2.35x; +tilelang adds ~26%. tilelang is GDN-only (Qwen3.5 family).
if not _ensure_flash_linear_attention_unconditional(eq):
logger.info(
"FLA install did not produce an importable runtime; skipping TileLang"
)
return False
if _model_wants_tilelang(model_name):
_ensure_tilelang_backend_unconditional(eq)
else:
logger.info(
"Model %r outside TileLang allowlist; FLA Triton path is sufficient",
model_name,
)
return True
def _fla_post_available(eq: Any) -> None:
# FLA already imports; repair tilelang if missing or on the broken tvm-ffi list.
if not _model_wants_tilelang(model_name):
return
if (
_installed_tvm_ffi_version() not in _TVM_FFI_BROKEN_VERSIONS
and _tilelang_importable()
):
return
_ensure_tilelang_backend_unconditional(eq)
def _causal_conv1d_install(eq: Any) -> bool:
ok = _install_package_wheel_first(
event_queue = eq,
import_name = "causal_conv1d",
display_name = "causal-conv1d",
pypi_name = "causal-conv1d",
pypi_version = _CAUSAL_CONV1D_PACKAGE_VERSION,
filename_prefix = "causal_conv1d",
release_tag = _CAUSAL_CONV1D_RELEASE_TAG,
release_base_url = (
"https://github.com/Dao-AILab/causal-conv1d/releases/download"
),
)
return bool(ok)
for gate_name, install_fn, post_fn in (
("is_flash_linear_attention_available", _fla_install, _fla_post_available),
("is_causal_conv1d_available", _causal_conv1d_install, None),
):
original = getattr(_iu, gate_name, None)
if original is None:
logger.info(
"%s missing on transformers.utils.import_utils; skipping hook",
gate_name,
)
continue
wrapped = _make_wrapper(original, install_fn, gate_name, post_fn)
setattr(_iu, gate_name, wrapped)
rebound = _rebind_in_already_imported_modules(
attr_name = gate_name, old_obj = original, new_obj = wrapped
)
logger.info(
"Installed fast-path hook on %s (rebound %d modules)", gate_name, rebound
)
def _should_try_runtime_flash_attn_install(max_seq_length: int) -> bool:
if os.getenv(_FLASH_ATTN_SKIP_ENV) == "1":
return False
if max_seq_length < _FLASH_ATTN_RUNTIME_MIN_SEQ_LEN:
return False
return sys.platform.startswith("linux")
def _ensure_flash_attn_for_long_context(event_queue: Any, max_seq_length: int) -> None:
if not _should_try_runtime_flash_attn_install(max_seq_length):
return
if has_blackwell_gpu():
_send_status(
event_queue,
"Skipping flash-attn install: Blackwell GPU detected (sm_100+); no compatible prebuilt wheel",
)
return
installed = _install_package_wheel_first(
event_queue = event_queue,
import_name = "flash_attn",
display_name = "flash-attn",
pypi_name = "flash-attn",
wheel_url_builder = flash_attn_wheel_url,
pypi_spec = "flash-attn",
pypi_status_message = "Installing flash-attn from PyPI for long-context training...",
)
if not installed:
_send_status(event_queue, "Continuing without flash-attn")
def _activate_transformers_version(model_name: str) -> None:
"""Activate the correct transformers version BEFORE any ML imports."""
# Ensure backend is on path for utils imports
backend_path = str(Path(__file__).resolve().parent.parent.parent)
if backend_path not in sys.path:
sys.path.insert(0, backend_path)
from utils.transformers_version import activate_transformers_for_subprocess
activate_transformers_for_subprocess(model_name)
def _adapt_for_mlx_vlm(items):
"""Adapt GPU-path VLM dataset output for mlx-vlm consumption.
The GPU path embeds PIL images inside messages content as
{"type": "image", "image": PIL_Image}. mlx-vlm's prepare_inputs
needs images at top-level to produce pixel_values — regardless of
model type. Extract them and leave bare {"type": "image"} placeholders.
"""
adapted = []
for item in items:
images = []
messages = []
for msg in item.get("messages", []):
content = msg.get("content", "")
if isinstance(content, list):
new_content = []
for part in content:
if isinstance(part, dict) and part.get("type") == "image":
img = part.get("image")
if img is not None:
images.append(img)
new_content.append({"type": "image"})
else:
new_content.append(part)
messages.append({"role": msg["role"], "content": new_content})
else:
messages.append(msg)
out = {"messages": messages}
if images:
out["image"] = images[0] if len(images) == 1 else images
elif "image" in item:
out["image"] = item["image"]
elif "images" in item:
out["images"] = item["images"]
adapted.append(out)
return adapted
_MLX_STUDIO_OPTIM_MAP = {
"adamw_8bit": "adamw",
"paged_adamw_8bit": "adamw",
"adamw_bnb_8bit": "adamw",
"paged_adamw_32bit": "adamw",
"adamw_torch": "adamw",
"adamw_torch_fused": "adamw",
"adamw": "adamw",
"adafactor": "adafactor",
"sgd": "sgd",
"adam": "adam",
"muon": "muon",
"lion": "lion",
}
_MLX_STUDIO_LR_SCHEDULERS = {"linear", "cosine", "constant"}
def _normalize_mlx_studio_optimizer(value):
raw = str(value or "adamw_8bit").strip().lower()
try:
return _MLX_STUDIO_OPTIM_MAP[raw]
except KeyError:
supported = ", ".join(sorted(_MLX_STUDIO_OPTIM_MAP))
raise ValueError(
f"Unsupported optimizer for MLX training: {value!r}. "
f"Supported values: {supported}."
)
def _normalize_mlx_studio_scheduler(value):
raw = str(value or "linear").strip().lower()
if raw not in _MLX_STUDIO_LR_SCHEDULERS:
supported = ", ".join(sorted(_MLX_STUDIO_LR_SCHEDULERS))
raise ValueError(
f"Unsupported LR scheduler for MLX training: {value!r}. "
f"Supported values: {supported}."
)
return raw
def _resolve_mlx_local_dataset_files(file_paths: list) -> list[str]:
"""Resolve Studio local dataset uploads without importing the GPU trainer."""
from utils.paths import resolve_dataset_path
all_files: list[str] = []
for dataset_file in file_paths or []:
file_path = (
dataset_file
if os.path.isabs(dataset_file)
else str(resolve_dataset_path(dataset_file))
)
file_path_obj = Path(file_path)
if file_path_obj.is_dir():
parquet_dir = (
file_path_obj / "parquet-files"
if (file_path_obj / "parquet-files").exists()
else file_path_obj
)
parquet_files = sorted(parquet_dir.glob("*.parquet"))
if parquet_files:
all_files.extend(str(p) for p in parquet_files)
continue
candidates: list[Path] = []
for ext in (".json", ".jsonl", ".csv", ".parquet"):
candidates.extend(sorted(file_path_obj.glob(f"*{ext}")))
if candidates:
all_files.extend(str(c) for c in candidates)
continue
raise ValueError(f"No supported data files in directory: {file_path_obj}")
all_files.append(str(file_path_obj))
return all_files
def _mlx_local_dataset_loader_for_files(files: list[str]) -> str:
first_ext = Path(files[0]).suffix.lower()
if first_ext in (".json", ".jsonl"):
return "json"
if first_ext == ".csv":
return "csv"
if first_ext == ".parquet":
return "parquet"
raise ValueError(f"Unsupported dataset format: {files[0]}")
def _run_mlx_training(event_queue, stop_queue, config):
"""Self-contained MLX training path for Apple Silicon.
Uses MLXTrainer from unsloth_zoo directly -- no torch/SFTTrainer needed.
Mirrors the event_queue protocol so the parent process pump works unchanged.
"""
import time
import gc
import math
import threading
import queue as _queue
from pathlib import Path
def _send(event_type, **kwargs):
if event_type == "status" and "message" not in kwargs:
sm = kwargs.get("status_message")
if sm is not None:
kwargs["message"] = sm
event_queue.put({"type": event_type, "ts": time.time(), **kwargs})
_send("status", status_message = "Loading MLX libraries...")
import mlx.core as mx
try:
from unsloth_zoo.mlx.loader import FastMLXModel
from unsloth_zoo.mlx.trainer import (
MLXTrainer,
MLXTrainingConfig,
train_on_responses_only,
)
except ImportError as e:
raise ImportError(
"Unsloth: MLX training requires unsloth-zoo with the MLX modules "
"(unsloth_zoo.mlx.loader / unsloth_zoo.mlx.trainer). Reinstall via "
"install.sh on Apple Silicon."
) from e
from datasets import load_dataset
if mx.metal.is_available():
info = mx.device_info()
rec_bytes = info.get("max_recommended_working_set_size", 0) or 0
if rec_bytes > 0:
memory_cap = int(rec_bytes * 0.85)
wired_cap = min(int(rec_bytes), memory_cap)
mx.set_memory_limit(memory_cap)
mx.set_wired_limit(wired_cap)
model_name = config["model_name"]
hf_token = config.get("hf_token") or None
if hf_token:
os.environ["HF_TOKEN"] = hf_token
if config.get("use_loftq"):
message = "LoftQ is not supported for MLX training yet."
_send("error", error = message)
raise NotImplementedError(message)
optim_name = _normalize_mlx_studio_optimizer(config.get("optim", "adamw_8bit"))
lr_scheduler_type = _normalize_mlx_studio_scheduler(
config.get("lr_scheduler_type", "linear")
)
# ── 1. Load model ──
# Force text-only if the dataset is not an image dataset, even if the model
# has vision capabilities (e.g. Qwen3.5-VL trained on plain alpaca text).
_send("status", status_message = f"Loading {model_name}...")
is_dataset_image = bool(config.get("is_dataset_image", False))
training_type = config.get("training_type", "LoRA/QLoRA")
use_lora = training_type == "LoRA/QLoRA"
model, tokenizer = FastMLXModel.from_pretrained(
model_name,
load_in_4bit = config.get("load_in_4bit", True),
full_finetuning = not use_lora,
text_only = None if is_dataset_image else True,
token = hf_token,
trust_remote_code = bool(config.get("trust_remote_code", False)),
random_state = config.get("random_seed", 3407),
)
is_vlm = bool(is_dataset_image and getattr(model, "_is_vlm_model", False))
model._is_vlm_model = is_vlm
# ── 2. Apply LoRA / full FT ──
# Pass gradient_checkpointing as string ("mlx"/"unsloth"/"none"/etc.)
# get_peft_model and MLXTrainer both accept strings and handle them.
gc_setting = config.get("gradient_checkpointing", "mlx")
if isinstance(gc_setting, str):
use_grad_checkpoint = (
gc_setting if gc_setting.lower() not in ("false", "") else False
)
else:
use_grad_checkpoint = gc_setting
if use_lora:
_send("status", status_message = "Configuring LoRA adapters...")
peft_kwargs = dict(
r = config.get("lora_r", 16),
lora_alpha = config.get("lora_alpha", 16),
lora_dropout = config.get("lora_dropout", 0.0),
use_rslora = config.get("use_rslora", False),
init_lora_weights = config.get("init_lora_weights", True),
random_state = config.get("random_seed", 3407),
target_modules = config.get("target_modules")
or [
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
use_gradient_checkpointing = use_grad_checkpoint,
)
finetune_language = config.get("finetune_language_layers", True)
finetune_attention = config.get("finetune_attention_modules", True)
finetune_mlp = config.get("finetune_mlp_modules", True)
finetune_vision = (
config.get("finetune_vision_layers", False) if is_vlm else False
)
if (
(finetune_attention or finetune_mlp)
and not finetune_language
and not finetune_vision
):
finetune_language = True
peft_kwargs["finetune_language_layers"] = finetune_language
peft_kwargs["finetune_attention_modules"] = finetune_attention
peft_kwargs["finetune_mlp_modules"] = finetune_mlp
if is_vlm:
peft_kwargs["finetune_vision_layers"] = finetune_vision
model = FastMLXModel.get_peft_model(model, **peft_kwargs)
# ── 3. Load dataset ──
_send("status", status_message = "Loading dataset...")
hf_dataset = config.get("hf_dataset", "")
subset = config.get("subset")
train_split = config.get("train_split", "train") or "train"
eval_split = config.get("eval_split")
slice_start = config.get("dataset_slice_start")
slice_end = config.get("dataset_slice_end")
def _slice(ds):
if slice_start is not None or slice_end is not None:
start = slice_start if slice_start is not None else 0
end = slice_end if slice_end is not None else len(ds) - 1
if end < start:
return ds.select([])
ds = ds.select(range(start, min(end + 1, len(ds))))
return ds
def _load_local(file_paths):
from datasets import load_from_disk
if len(file_paths) == 1:
p = Path(file_paths[0])
if p.is_dir() and (
(p / "dataset_info.json").exists() or (p / "state.json").exists()
):
return load_from_disk(str(p))
all_files = _resolve_mlx_local_dataset_files(file_paths)
if not all_files:
raise ValueError("No local dataset files found")
loader = _mlx_local_dataset_loader_for_files(all_files)
return load_dataset(loader, data_files = all_files, split = "train")
if hf_dataset:
load_kwargs = {"split": train_split, "token": hf_token}
if subset:
load_kwargs["name"] = subset
dataset = load_dataset(hf_dataset, **load_kwargs)
dataset = _slice(dataset)
elif config.get("local_datasets"):
dataset = _load_local(config["local_datasets"])
dataset = _slice(dataset)
else:
raise ValueError("No dataset specified")
# Eval dataset (separate split or local file)
eval_dataset = None
if eval_split and hf_dataset:
eval_kwargs = {"split": eval_split, "token": hf_token}
if subset:
eval_kwargs["name"] = subset
try:
eval_dataset = load_dataset(hf_dataset, **eval_kwargs)
except Exception as e:
_send("status", status_message = f"Eval split load failed: {e}")
eval_dataset = None
elif config.get("local_eval_datasets"):
eval_dataset = _load_local(config["local_eval_datasets"])
# ── 3b. Format dataset (VLM or text) ──
# Reuse the GPU path's format pipeline for both VLM (auto-detects OCR/caption/
# llava/sharegpt+images) and text (alpaca/sharegpt/chatml → "text" column).
format_type = config.get("format_type", "")
try:
from utils.datasets import format_and_template_dataset
def _fmt_progress(status_message = "", **_kw):
_send("status", status_message = status_message)
if is_vlm:
_send("status", status_message = "Formatting VLM dataset...")
vlm_info = format_and_template_dataset(
dataset,
model_name = model_name,
tokenizer = tokenizer,
is_vlm = True,
dataset_name = hf_dataset or "local",
progress_callback = _fmt_progress,
)
if vlm_info.get("success"):
dataset = _adapt_for_mlx_vlm(vlm_info["dataset"])
else:
errors = vlm_info.get("errors", [])
raise ValueError(
f"VLM dataset format conversion failed: {'; '.join(errors)}"
)
if eval_dataset is not None:
ev_info = format_and_template_dataset(
eval_dataset,
model_name = model_name,
tokenizer = tokenizer,
is_vlm = True,
dataset_name = hf_dataset or "local",
)
if ev_info.get("success"):
eval_dataset = _adapt_for_mlx_vlm(ev_info["dataset"])
elif format_type:
_send("status", status_message = f"Formatting dataset ({format_type})...")
info = format_and_template_dataset(
dataset,
model_name = model_name,
tokenizer = tokenizer,
is_vlm = False,
format_type = format_type,
dataset_name = hf_dataset or "local",
)
if info.get("success", True):
dataset = info.get("dataset", dataset)
if eval_dataset is not None:
ev = format_and_template_dataset(
eval_dataset,
model_name = model_name,
tokenizer = tokenizer,
is_vlm = False,
format_type = format_type,
dataset_name = hf_dataset or "local",
)
if ev.get("success", True):
eval_dataset = ev.get("dataset", eval_dataset)
except ImportError:
_send("status", status_message = "Format helper unavailable, using raw dataset")
# ── 4. Resolve training steps ──
max_steps = config.get("max_steps", 0) or 0
num_epochs = config.get("num_epochs", 3)
max_seq_length = config.get("max_seq_length", 2048)
batch_size = config.get("batch_size", 4)
grad_accum = config.get("gradient_accumulation_steps", 4)
if max_steps <= 0:
max_steps = max(
1,
math.ceil(len(dataset) / batch_size / grad_accum) * num_epochs,
)
lr_value = float(config.get("learning_rate", "2e-4"))
# Warmup: prefer warmup_steps; fall back to warmup_ratio
warmup_steps = config.get("warmup_steps")
warmup_ratio = config.get("warmup_ratio")
if warmup_steps is None and warmup_ratio is not None:
warmup_steps = int(round(warmup_ratio * max_steps))
if warmup_steps is None:
warmup_steps = 5
# ── 5. Build output dir ──
output_dir = config.get("output_dir", "")
if not output_dir:
output_dir = f"{model_name.replace('/', '_')}_{int(time.time())}"
# Resolve to ~/.unsloth/studio/outputs/ so the export page can find it
from utils.paths import resolve_output_dir, ensure_dir
output_dir = str(resolve_output_dir(output_dir))
ensure_dir(Path(output_dir))
# ── 6. Create trainer ──
eval_steps_val = config.get("eval_steps", 0) or 0
if isinstance(eval_steps_val, float) and 0 < eval_steps_val < 1:
# Studio sometimes sends fraction-of-total-steps
eval_steps_val = max(1, int(eval_steps_val * max_steps))
else:
eval_steps_val = int(eval_steps_val)
# MLX: per-element clip to [-1, 1]; norm clip disabled (it needs a
# global reduction that breaks MLX's eager pipeline). 1.0 (not 5.0):
# |g_i| > 5 rarely fires, so the historical 5.0 was effectively no-op.
max_grad_norm = 0.0
max_grad_value = 1.0 # TODO: expose MLX grad-clip in Studio UI for power users
trainer = MLXTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
eval_dataset = eval_dataset,
args = MLXTrainingConfig(
per_device_train_batch_size = batch_size,
gradient_accumulation_steps = grad_accum,
max_steps = max_steps,
learning_rate = lr_value,
warmup_steps = warmup_steps,
lr_scheduler_type = lr_scheduler_type,
optim = optim_name,
weight_decay = float(config.get("weight_decay", 0.001) or 0.001),
max_grad_norm = max_grad_norm,
max_grad_value = max_grad_value,
logging_steps = 1,
max_seq_length = max_seq_length,
seed = config.get("random_seed", 3407),
use_cce = True,
compile = True,
gradient_checkpointing = use_grad_checkpoint,
streaming = is_vlm,
packing = bool(config.get("packing", False)),
output_dir = output_dir,
save_steps = int(config.get("save_steps", 0) or 0),
eval_steps = eval_steps_val,
),
)
# Tell the parent that eval is configured so the frontend shows the eval chart
if eval_dataset is not None and eval_steps_val > 0:
_send("eval_configured")
# ── 7. Apply train_on_responses_only if requested ──
if config.get("train_on_completions", False):
_send("status", status_message = "Configuring response-only training...")
try:
from utils.datasets import (
MODEL_TO_TEMPLATE_MAPPER,
TEMPLATE_TO_RESPONSES_MAPPER,
)
template_name = MODEL_TO_TEMPLATE_MAPPER.get(model_name.lower())
markers = (
TEMPLATE_TO_RESPONSES_MAPPER.get(template_name)
if template_name
else None
)
if markers:
trainer = train_on_responses_only(
trainer,
instruction_part = markers["instruction"],
response_part = markers["response"],
)
else:
_send(
"status",
status_message = f"train_on_completions skipped (no template for {model_name})",
)
except Exception as e:
_send("status", status_message = f"train_on_completions failed: {e}")
# ── 8. Setup wandb / tensorboard ──
wandb_run = None
tb_writer = None
if config.get("enable_wandb", False):
try:
import wandb as _wandb
wandb_token = config.get("wandb_token")
if wandb_token:
os.environ["WANDB_API_KEY"] = wandb_token
_wandb_sensitive = {"hf_token", "wandb_token"}
wandb_run = _wandb.init(
project = config.get("wandb_project") or "unsloth-mlx",
config = {k: v for k, v in config.items() if k not in _wandb_sensitive},
reinit = True,
)
except Exception as e:
_send("status", status_message = f"wandb init failed: {e}")
if config.get("enable_tensorboard", False):
try:
from tensorboardX import SummaryWriter
except ImportError:
try:
from torch.utils.tensorboard import SummaryWriter
except ImportError:
SummaryWriter = None
if SummaryWriter is not None:
try:
tb_dir = config.get("tensorboard_dir") or f"{output_dir}/runs"
tb_writer = SummaryWriter(log_dir = tb_dir)
except Exception as e:
_send("status", status_message = f"tensorboard init failed: {e}")
else:
_send(
"status",
status_message = "tensorboard unavailable (install tensorboardX)",
)
# ── 9. Real-time progress callback ──
_send("status", status_message = f"Training {model_name}...")
def _on_step(
step,
total,
loss,
lr,
tok_s,
peak_gb,
elapsed,
num_tokens,
grad_norm = None,
):
eta = (elapsed / step * (total - step)) if step > 0 else 0
_send(
"progress",
step = step,
epoch = round(step / total * num_epochs, 2) if total > 0 else 0,
loss = loss,
learning_rate = lr,
total_steps = total,
elapsed_seconds = elapsed,
eta_seconds = max(0, eta),
grad_norm = grad_norm,
num_tokens = num_tokens,
eval_loss = None,
status_message = None,
peak_memory_gb = peak_gb,
)
if wandb_run is not None:
try:
wandb_run.log(
{
"train/loss": loss,
"train/learning_rate": lr,
"train/tokens_per_sec": tok_s,
"train/peak_gb": peak_gb,
"train/num_tokens": num_tokens,
**(
{"train/grad_norm": grad_norm}
if grad_norm is not None
else {}
),
},
step = step,
)
except Exception:
pass
if tb_writer is not None:
try:
tb_writer.add_scalar("train/loss", loss, step)
tb_writer.add_scalar("train/learning_rate", lr, step)
tb_writer.add_scalar("train/tokens_per_sec", tok_s, step)
tb_writer.add_scalar("train/peak_gb", peak_gb, step)
if grad_norm is not None:
tb_writer.add_scalar("train/grad_norm", grad_norm, step)
except Exception:
pass
trainer.add_step_callback(_on_step)
def _on_eval(step, eval_loss, perplexity):
_send("progress", step = step, eval_loss = eval_loss)
if wandb_run is not None:
try:
wandb_run.log(
{"eval/loss": eval_loss, "eval/perplexity": perplexity}, step = step
)
except Exception:
pass
if tb_writer is not None:
try:
tb_writer.add_scalar("eval/loss", eval_loss, step)
tb_writer.add_scalar("eval/perplexity", perplexity, step)
except Exception:
pass
trainer.add_eval_callback(_on_eval)
# ── 10. Stop signal polling ──
_stop_save = [True] # mutable so thread can update; [save_flag]
def _poll_stop():
while True:
try:
msg = stop_queue.get(timeout = 1.0)
if msg and msg.get("type") == "stop":
_stop_save[0] = msg.get("save", True)
trainer.stop_requested = True
return
except _queue.Empty:
continue
except (EOFError, OSError):
# why safe: pipe permanently broken, no further messages can arrive
return
stop_thread = threading.Thread(target = _poll_stop, daemon = True)
stop_thread.start()
# ── 11. Run training ──
gc.collect()
mx.synchronize()
trainer.train()
# ── 12. Save and finalize ──
if trainer.stop_requested and not _stop_save[0]:
# User clicked "Cancel" (save=False) — skip saving
_send("complete", output_dir = None, status_message = "Training cancelled")
else:
_send("status", status_message = "Saving model...")
mx.synchronize()
trainer.save_model(output_dir)
_send("complete", output_dir = output_dir, status_message = "Training completed")
if tb_writer is not None:
try:
tb_writer.close()
except Exception:
pass
if wandb_run is not None:
try:
wandb_run.finish()
except Exception:
pass
def run_training_process(
*,
event_queue: Any,
stop_queue: Any,
config: dict,
) -> None:
"""Subprocess entrypoint. Fresh Python — no stale module state.
Args:
event_queue: mp.Queue for sending progress/status/error events to parent.
stop_queue: mp.Queue for receiving stop commands from parent.
config: Training configuration dict with all parameters.
"""
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["PYTHONWARNINGS"] = (
"ignore" # Suppress warnings at C-level before imports
)
# Offline auto-detect: skip ~25s of HF retries per call when DNS is
# dead. Scoped to this subprocess (orchestrator spawns a fresh one).
if "HF_HUB_OFFLINE" not in os.environ:
import socket as _socket
import threading as _threading
# Daemon thread so we don't mutate process-wide setdefaulttimeout.
_result: list = [None]
def _probe() -> None:
try:
_socket.gethostbyname("huggingface.co")
_result[0] = False
except Exception:
_result[0] = True
_t = _threading.Thread(target = _probe, daemon = True)
_t.start()
_t.join(2.0)
if _result[0] is None or _result[0] is True:
os.environ["HF_HUB_OFFLINE"] = "1"
os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
os.environ.setdefault("HF_DATASETS_OFFLINE", "1")
# logger isn't configured yet; print to stderr instead.
print(
"huggingface.co unreachable; HF_HUB_OFFLINE=1 set for this worker.",
file = sys.stderr,
flush = True,
)
import warnings
from loggers.config import LogConfig
if os.getenv("ENVIRONMENT_TYPE", "production") == "production":
warnings.filterwarnings("ignore")
LogConfig.setup_logging(
service_name = "unsloth-studio-training-worker",
env = os.getenv("ENVIRONMENT_TYPE", "production"),
)
apply_gpu_ids(config.get("resolved_gpu_ids"))
model_name = config["model_name"]
# ── 0. MLX FAST-PATH (must run before any torch/transformers imports) ──
# Apple Silicon uses MLXTrainer directly -- skip transformers version
# activation, causal-conv1d install, and torch imports entirely.
backend_path = str(Path(__file__).resolve().parent.parent.parent)
if backend_path not in sys.path:
sys.path.insert(0, backend_path)
from utils.hardware import hardware as _hw
_hw.detect_hardware()
if _hw.DEVICE == _hw.DeviceType.MLX:
if config.get("is_dataset_audio"):
event_queue.put(
{
"type": "error",
"error": "Audio dataset training is not yet supported on Apple Silicon.",
"stack": "",
"ts": time.time(),
}
)
return
# Activate correct transformers version (Gemma-4 needs 5.5.0, etc.)
# Must happen before any transformers/mlx-lm imports in _run_mlx_training.
try:
_activate_transformers_version(model_name)
except Exception:
pass # Non-fatal: fall through with whatever version is installed
try:
_run_mlx_training(event_queue, stop_queue, config)
except Exception as exc:
event_queue.put(
{
"type": "error",
"error": str(exc),
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 1. Activate correct transformers version BEFORE any ML imports ──
try:
_activate_transformers_version(model_name)
except Exception as exc:
event_queue.put(
{
"type": "error",
"error": f"Failed to activate transformers version: {exc}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 1a. Auto-enable trust_remote_code for NemotronH/Nano models ──
# NemotronH has config parsing bugs in transformers that require
# trust_remote_code=True as a workaround. Other transformers 5.x models
# (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
# bypasses the compiler (disabling fused CE).
# NOTE: Must NOT match Llama-Nemotron (standard Llama architecture).
_NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
_lowered = model_name.lower()
if (
any(sub in _lowered for sub in _NEMOTRON_TRUST_SUBSTRINGS)
and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
and not config.get("trust_remote_code", False)
):
config["trust_remote_code"] = True
logger.info(
"Auto-enabled trust_remote_code for Nemotron model: %s",
model_name,
)
# ── 1b. Install fast-path kernel libraries for the chosen model.
#
# 1) causal-conv1d ALWAYS runs eagerly via the substring path.
# Some SSM modeling files (nemotron_h, falcon_h1, granitemoehybrid)
# use `lazy_load_kernel("causal-conv1d")` directly and never call
# transformers' `is_causal_conv1d_available()`, so the runtime
# hook on that gate would not fire for them.
# 2) FLA + tilelang: primary gate is the runtime hook on transformers'
# `is_flash_linear_attention_available`. Models whose architecture
# queries that gate auto-trigger the install; others never pay.
# `_install_fast_path_hooks` also wraps `is_causal_conv1d_available`
# as a defence in depth for newer modeling files that do use it.
# 3) mamba-ssm + flash-attn keep their existing substring / size gates.
# 4) `UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS=1` falls back to the
# substring path for FLA / tilelang.
try:
_ensure_causal_conv1d_fast_path(event_queue, model_name)
if os.getenv(_FAST_PATH_HOOKS_SKIP_ENV) == "1":
_ensure_flash_linear_attention(event_queue, model_name)
_ensure_tilelang_backend(event_queue, model_name)
else:
_install_fast_path_hooks(event_queue, model_name)
_ensure_mamba_ssm(event_queue, model_name)
_ensure_flash_attn_for_long_context(
event_queue,
int(config.get("max_seq_length", 2048)),
)
except Exception as exc:
event_queue.put(
{
"type": "error",
"error": (
f"Please choose another model to train, since "
f"a fast-path kernel library "
f"(causal-conv1d / flash-linear-attention / "
f"mamba-ssm / tilelang) failed to install "
f"with error: {exc}"
),
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 1c. Set fork start method so dataset.map() can multiprocess ──
# The parent launched us via spawn (clean process), but the compiled
# SFTTrainer checks get_start_method() and disables num_proc if not "fork".
# Linux only: fork is the default start method and is safe here (no CUDA
# context exists yet). macOS defaults to spawn since Python 3.8 because
# fork is unsafe with macOS frameworks (Metal/MPS, CoreFoundation) --
# do NOT override on macOS. Windows has no fork at all.
if sys.platform == "linux":
import multiprocessing as _mp
try:
_mp.set_start_method("fork", force = True)
except RuntimeError:
pass # Already set
# ── 1c. On Windows, check Triton availability (must be before import torch) ──
if sys.platform == "win32":
try:
import triton # noqa: F401
logger.info("Triton available — torch.compile enabled")
except ImportError:
os.environ["TORCHDYNAMO_DISABLE"] = "1"
logger.warning(
"Triton not found on Windows — torch.compile disabled. "
'Install for better performance: pip install "triton-windows<3.7"'
)
# ── 2. Now import ML libraries (fresh in this clean process) ──
try:
_send_status(event_queue, "Importing Unsloth...")
backend_path = str(Path(__file__).resolve().parent.parent.parent)
if backend_path not in sys.path:
sys.path.insert(0, backend_path)
from core.training.trainer import UnslothTrainer, TrainingProgress
from utils.paths import (
ensure_dir,
resolve_output_dir,
resolve_tensorboard_dir,
datasets_root,
)
import transformers
logger.info("Subprocess loaded transformers %s", transformers.__version__)
except Exception as exc:
event_queue.put(
{
"type": "error",
"error": f"Failed to import ML libraries: {exc}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 2b. EMBEDDING MODEL FAST-PATH ──
# Embedding models use a completely different pipeline (FastSentenceTransformer
# + SentenceTransformerTrainer + MultipleNegativesRankingLoss) so we branch
# early and handle the entire flow in a self-contained function.
if config.get("is_embedding", False):
try:
_run_embedding_training(event_queue, stop_queue, config)
except Exception as exc:
event_queue.put(
{
"type": "error",
"error": str(exc),
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 3. Create a fresh trainer instance ──
trainer = UnslothTrainer()
# Wire up progress callback → event_queue
def _on_progress(progress: TrainingProgress):
has_train_loss = progress.step > 0 and progress.loss is not None
has_eval_loss = progress.eval_loss is not None
if has_train_loss or has_eval_loss:
event_queue.put(
{
"type": "progress",
"step": progress.step,
"epoch": progress.epoch,
"loss": progress.loss,
"learning_rate": progress.learning_rate,
"total_steps": progress.total_steps,
"elapsed_seconds": progress.elapsed_seconds,
"eta_seconds": progress.eta_seconds,
"grad_norm": progress.grad_norm,
"num_tokens": progress.num_tokens,
"eval_loss": progress.eval_loss,
"status_message": progress.status_message,
"ts": time.time(),
}
)
if progress.status_message:
_send_status(event_queue, progress.status_message)
trainer.add_progress_callback(_on_progress)
# Wire up stop_queue polling to trainer.should_stop
import threading
import queue as _queue
def _poll_stop():
while True:
try:
msg = stop_queue.get(timeout = 1.0)
if msg and msg.get("type") == "stop":
save = msg.get("save", True)
trainer.should_stop = True
trainer.save_on_stop = save
logger.info("Stop signal received (save=%s)", save)
return
except _queue.Empty:
continue
except (EOFError, OSError):
return
stop_thread = threading.Thread(target = _poll_stop, daemon = True)
stop_thread.start()
# ── 4. Execute the training pipeline ──
# Order: detect → dataset → model → prepare → train
# Dataset processing (including LLM-assisted detection) runs BEFORE model
# loading so both never occupy VRAM at the same time.
try:
hf_token = config.get("hf_token", "")
hf_token = hf_token if hf_token and hf_token.strip() else None
# ── 4a. Lightweight detection + tokenizer (no VRAM) ──
_send_status(event_queue, "Detecting model type...")
trainer.pre_detect_and_load_tokenizer(
model_name = model_name,
max_seq_length = config["max_seq_length"],
hf_token = hf_token,
is_dataset_image = config.get("is_dataset_image", False),
is_dataset_audio = config.get("is_dataset_audio", False),
trust_remote_code = config.get("trust_remote_code", False),
)
if trainer.should_stop:
event_queue.put({"type": "complete", "output_dir": None, "ts": time.time()})
return
# ── 4b. Load and format dataset (LLM helper may use VRAM briefly) ──
_send_status(event_queue, "Loading and formatting dataset...")
hf_dataset = config.get("hf_dataset", "")
training_type = config.get("training_type", "LoRA/QLoRA")
_is_cpt_for_dataset = training_type == "Continued Pretraining"
dataset_result = trainer.load_and_format_dataset(
dataset_source = hf_dataset if hf_dataset and hf_dataset.strip() else None,
format_type = config.get("format_type", ""),
local_datasets = config.get("local_datasets") or None,
local_eval_datasets = config.get("local_eval_datasets") or None,
custom_format_mapping = config.get("custom_format_mapping"),
subset = config.get("subset"),
train_split = config.get("train_split", "train"),
eval_split = config.get("eval_split"),
eval_steps = config.get("eval_steps", 0.00),
dataset_slice_start = config.get("dataset_slice_start"),
dataset_slice_end = config.get("dataset_slice_end"),
is_cpt = _is_cpt_for_dataset,
)
if isinstance(dataset_result, tuple):
dataset, eval_dataset = dataset_result
else:
dataset = dataset_result
eval_dataset = None
# [DEBUG] Print first sample before model is loaded
# dataset is a dict {"dataset": <Dataset>, "detected_format": ..., ...}
# or a raw Dataset for audio paths
# try:
# ds = dataset["dataset"] if isinstance(dataset, dict) else dataset
# print(
# f"\n[DEBUG] Dataset loaded BEFORE model. type={type(ds).__name__}, len={len(ds)}",
# flush = True,
# )
# print(f"[DEBUG] Columns: {ds.column_names}", flush = True)
# sample = ds[0]
# preview = {k: str(v)[:300] for k, v in sample.items()}
# print(f"[DEBUG] First sample: {preview}\n", flush = True)
# except Exception as e:
# print(
# f"[DEBUG] Could not preview first sample: {type(e).__name__}: {e}",
# flush = True,
# )
# Disable eval if eval_steps <= 0
eval_steps = config.get("eval_steps", 0.00)
if eval_steps is not None and float(eval_steps) <= 0:
eval_dataset = None
# Tell the parent process that eval is configured so the frontend
# shows "Waiting for first evaluation step..." instead of "not configured"
if eval_dataset is not None:
event_queue.put(
{
"type": "eval_configured",
"ts": time.time(),
}
)
if dataset is None or trainer.should_stop:
if trainer.should_stop:
event_queue.put(
{"type": "complete", "output_dir": None, "ts": time.time()}
)
else:
event_queue.put(
{
"type": "error",
"error": trainer.training_progress.error
or "Failed to load dataset",
"stack": "",
"ts": time.time(),
}
)
return
# ── Start tqdm monitor early so it captures download + tokenization bars ──
import threading as _th
_tqdm_stop = _th.Event()
def _monitor_tqdm():
from tqdm.auto import tqdm as _tqdm_cls
while not _tqdm_stop.is_set():
for bar in list(getattr(_tqdm_cls, "_instances", set())):
try:
n, total = bar.n or 0, bar.total or 0
desc = getattr(bar, "desc", "") or ""
if total > 0 and n > 0 and desc:
pct = min(int(n * 100 / total), 100)
_send_status(
event_queue, f"{desc.strip()} {pct}% ({n:,}/{total:,})"
)
except (AttributeError, ReferenceError):
pass
_tqdm_stop.wait(3)
_tqdm_thread = _th.Thread(target = _monitor_tqdm, daemon = True)
_tqdm_thread.start()
training_type = config.get("training_type", "LoRA/QLoRA")
is_cpt = training_type == "Continued Pretraining"
use_lora = training_type in ("LoRA/QLoRA", "Continued Pretraining")
cpt_trains_embeddings = False
# ── 4c. Load training model (uses VRAM — dataset already formatted) ──
_send_status(event_queue, "Loading model...")
success = trainer.load_model(
model_name = model_name,
max_seq_length = config["max_seq_length"],
load_in_4bit = config["load_in_4bit"],
full_finetuning = not use_lora,
hf_token = hf_token,
is_dataset_image = config.get("is_dataset_image", False),
is_dataset_audio = config.get("is_dataset_audio", False),
trust_remote_code = config.get("trust_remote_code", False),
gpu_ids = config.get("resolved_gpu_ids"),
)
if not success or trainer.should_stop:
if trainer.should_stop:
event_queue.put(
{"type": "complete", "output_dir": None, "ts": time.time()}
)
else:
error_msg = trainer.training_progress.error or "Failed to load model"
event_queue.put(
{
"type": "error",
"error": error_msg,
"stack": "",
"ts": time.time(),
}
)
return
# ── 4d. Prepare model (LoRA, full finetuning, or CPT) ──
if is_cpt:
_send_status(event_queue, "Configuring LoRA for continued pretraining...")
# embed_tokens (if the user included it) goes to modules_to_save —
# trained full-precision at embedding_learning_rate. lm_head stays as
# a LoRA target for merge compatibility (see unsloth PR #4106).
_user_modules = config.get("target_modules") or []
wants_embed = "embed_tokens" in _user_modules
cpt_trains_embeddings = wants_embed
cpt_target_modules = [m for m in _user_modules if m != "embed_tokens"]
if not cpt_target_modules:
cpt_target_modules = [
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
"lm_head",
]
success = trainer.prepare_model_for_training(
use_lora = True,
target_modules = cpt_target_modules,
modules_to_save = ["embed_tokens"] if wants_embed else None,
lora_r = config.get("lora_r", 128),
lora_alpha = config.get("lora_alpha", 32),
lora_dropout = config.get("lora_dropout", 0.0),
use_gradient_checkpointing = config.get(
"gradient_checkpointing", "unsloth"
),
use_rslora = config.get("use_rslora", False),
use_loftq = config.get("use_loftq", False),
)
elif use_lora:
_send_status(event_queue, "Configuring LoRA adapters...")
success = trainer.prepare_model_for_training(
use_lora = True,
finetune_vision_layers = config.get("finetune_vision_layers", True),
finetune_language_layers = config.get("finetune_language_layers", True),
finetune_attention_modules = config.get(
"finetune_attention_modules", True
),
finetune_mlp_modules = config.get("finetune_mlp_modules", True),
target_modules = config.get("target_modules"),
lora_r = config.get("lora_r", 16),
lora_alpha = config.get("lora_alpha", 16),
lora_dropout = config.get("lora_dropout", 0.0),
use_gradient_checkpointing = config.get(
"gradient_checkpointing", "unsloth"
),
use_rslora = config.get("use_rslora", False),
use_loftq = config.get("use_loftq", False),
)
else:
_send_status(event_queue, "Preparing model for full finetuning...")
success = trainer.prepare_model_for_training(use_lora = False)
if not success or trainer.should_stop:
if trainer.should_stop:
event_queue.put(
{"type": "complete", "output_dir": None, "ts": time.time()}
)
else:
event_queue.put(
{
"type": "error",
"error": trainer.training_progress.error
or "Failed to prepare model",
"stack": "",
"ts": time.time(),
}
)
return
lr_default = "5e-5" if is_cpt else "2e-4"
try:
lr_value = float(config.get("learning_rate", lr_default))
except ValueError:
event_queue.put(
{
"type": "error",
"error": f"Invalid learning rate: {config.get('learning_rate')}",
"stack": "",
"ts": time.time(),
}
)
return
# embedding_learning_rate is validated by the Pydantic model (Optional[float],
# gt=0, lt=1.0); if present it is already a finite float in range.
embedding_lr_value = config.get("embedding_learning_rate")
if is_cpt:
if cpt_trains_embeddings:
if embedding_lr_value is None:
# Default embedding_learning_rate = lr/10 per Unsloth's CPT notebook.
embedding_lr_value = lr_value / 10.0
logger.info(
f"CPT: using default embedding_learning_rate={embedding_lr_value:.1e} "
f"(lr/10). Set explicitly to override.\n"
)
elif embedding_lr_value is not None:
logger.warning(
"CPT: embedding_learning_rate was provided but embed_tokens is "
"not being trained; ignoring the override.\n"
)
embedding_lr_value = None
# Generate output dir
resume_from_checkpoint = config.get("resume_from_checkpoint")
output_dir = config.get("output_dir") or _output_dir_from_resume_checkpoint(
resume_from_checkpoint
)
if not output_dir:
output_dir = f"{model_name.replace('/', '_')}_{int(time.time())}"
output_dir = str(resolve_output_dir(output_dir))
ensure_dir(Path(output_dir))
tensorboard_dir = config.get("tensorboard_dir")
if config.get("enable_tensorboard", False):
tensorboard_dir = str(resolve_tensorboard_dir(tensorboard_dir))
ensure_dir(Path(tensorboard_dir))
# Start training (directly — no inner thread, we ARE the subprocess)
dataset_display = (
config.get("hf_dataset", "") or config.get("uploaded_file", "") or ""
)
_send_status(
event_queue,
f'Training "{model_name}"'
+ (f"\nDataset = {dataset_display}" if dataset_display else ""),
)
max_steps = config.get("max_steps", 0)
save_steps = config.get("save_steps", 0)
trainer._train_worker(
dataset,
output_dir = output_dir,
num_epochs = config.get("num_epochs", 3),
learning_rate = lr_value,
embedding_learning_rate = embedding_lr_value,
batch_size = config.get("batch_size", 2),
gradient_accumulation_steps = config.get("gradient_accumulation_steps", 4),
warmup_steps = config.get("warmup_steps"),
warmup_ratio = config.get("warmup_ratio"),
max_steps = max_steps if max_steps and max_steps > 0 else 0,
save_steps = save_steps if save_steps and save_steps > 0 else 0,
weight_decay = config.get("weight_decay", 0.001),
random_seed = config.get("random_seed", 3407),
packing = config.get("packing", False),
train_on_completions = False
if is_cpt
else config.get("train_on_completions", False),
enable_wandb = config.get("enable_wandb", False),
wandb_project = config.get("wandb_project", "unsloth-training"),
wandb_token = config.get("wandb_token"),
enable_tensorboard = config.get("enable_tensorboard", False),
tensorboard_dir = tensorboard_dir,
eval_dataset = eval_dataset,
eval_steps = eval_steps,
max_seq_length = config.get("max_seq_length", 2048),
optim = config.get("optim", "adamw_8bit"),
lr_scheduler_type = config.get("lr_scheduler_type", "linear"),
is_cpt = is_cpt,
resume_from_checkpoint = resume_from_checkpoint,
)
_tqdm_stop.set()
# Check final state
progress = trainer.get_training_progress()
if progress.error:
event_queue.put(
{
"type": "error",
"error": progress.error,
"stack": "",
"ts": time.time(),
}
)
else:
saved_output_dir = (
None if trainer.should_stop and not trainer.save_on_stop else output_dir
)
event_queue.put(
{
"type": "complete",
"output_dir": saved_output_dir,
"status_message": progress.status_message or "Training completed",
"ts": time.time(),
}
)
except Exception as exc:
event_queue.put(
{
"type": "error",
"error": str(exc),
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
def _send_status(event_queue: Any, message: str) -> None:
"""Send a status update to the parent process."""
event_queue.put(
{
"type": "status",
"message": message,
"ts": time.time(),
}
)
def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) -> None:
"""Self-contained embedding model training pipeline.
Uses FastSentenceTransformer + SentenceTransformerTrainer +
MultipleNegativesRankingLoss — completely separate from the
LLM/VLM/audio paths in UnslothTrainer.
Mirrors the pattern from the reference embedding notebooks:
All_MiniLM_L6_v2.py, BGE_M3.py, EmbeddingGemma_300M.py,
ModernBert.py, Qwen3_Embedding_0_6B.py
"""
import math
import queue as _queue
import threading
model_name = config["model_name"]
training_start_time = time.time()
# ── 1. Import embedding-specific libraries ──
_send_status(event_queue, "Importing embedding libraries...")
try:
from unsloth import FastSentenceTransformer, is_bfloat16_supported
from sentence_transformers import (
SentenceTransformerTrainer,
SentenceTransformerTrainingArguments,
)
from sentence_transformers.losses import MultipleNegativesRankingLoss
from sentence_transformers.training_args import BatchSamplers
from datasets import load_dataset, Dataset
from transformers import TrainerCallback
from utils.paths import datasets_root, resolve_output_dir
except ImportError as e:
event_queue.put(
{
"type": "error",
"error": f"Failed to import embedding libraries: {e}. "
"Ensure 'sentence_transformers' and 'unsloth' are installed.",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── Stop signal handling ──
_should_stop = False
_save_on_stop = True
def _poll_stop():
nonlocal _should_stop, _save_on_stop
while True:
try:
msg = stop_queue.get(timeout = 1.0)
if msg and msg.get("type") == "stop":
_save_on_stop = msg.get("save", True)
_should_stop = True
logger.info(
"Embedding training: stop signal received (save=%s)",
_save_on_stop,
)
return
except _queue.Empty:
continue
except (EOFError, OSError):
return
stop_thread = threading.Thread(target = _poll_stop, daemon = True)
stop_thread.start()
# ── 2. Load model ──
_send_status(event_queue, "Loading embedding model...")
try:
hf_token = config.get("hf_token", "")
hf_token = hf_token if hf_token and hf_token.strip() else None
max_seq_length = config.get("max_seq_length", 512)
training_type = config.get("training_type", "LoRA/QLoRA")
use_lora = training_type == "LoRA/QLoRA"
model = FastSentenceTransformer.from_pretrained(
model_name = model_name,
max_seq_length = max_seq_length,
full_finetuning = not use_lora,
token = hf_token,
)
except Exception as e:
event_queue.put(
{
"type": "error",
"error": f"Failed to load embedding model '{model_name}': {e}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
if _should_stop:
event_queue.put({"type": "complete", "output_dir": None, "ts": time.time()})
return
# ── 3. Apply LoRA ──
if use_lora:
_send_status(event_queue, "Configuring LoRA adapters (FEATURE_EXTRACTION)...")
try:
gradient_checkpointing = config.get("gradient_checkpointing", False)
# Normalize: "none" or empty → False
if gradient_checkpointing in ("none", "", None):
gradient_checkpointing = False
model = FastSentenceTransformer.get_peft_model(
model,
r = config.get("lora_r", 32),
target_modules = config.get("target_modules")
or ["q_proj", "k_proj", "v_proj", "o_proj"],
lora_alpha = config.get("lora_alpha", 64),
lora_dropout = config.get("lora_dropout", 0.0),
bias = "none",
use_gradient_checkpointing = gradient_checkpointing,
random_state = config.get("random_seed", 3407),
use_rslora = config.get("use_rslora", False),
loftq_config = {"loftq_bits": 4, "loftq_iter": 1}
if config.get("use_loftq")
else None,
task_type = "FEATURE_EXTRACTION",
)
except Exception as e:
event_queue.put(
{
"type": "error",
"error": f"Failed to configure LoRA for embedding model: {e}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
if _should_stop:
event_queue.put({"type": "complete", "output_dir": None, "ts": time.time()})
return
# ── 4. Load dataset ──
_send_status(event_queue, "Loading dataset...")
try:
hf_dataset = config.get("hf_dataset", "")
local_datasets = config.get("local_datasets") or []
subset = config.get("subset") or None
train_split = config.get("train_split", "train") or "train"
if hf_dataset and hf_dataset.strip():
hf_token = config.get("hf_token", "")
hf_token = hf_token if hf_token and hf_token.strip() else None
dataset = load_dataset(
hf_dataset.strip(),
subset,
split = train_split,
token = hf_token,
)
elif local_datasets:
# Load from local file(s) — mirrors the non-embedding pipeline's
# directory handling so recipe outputs (parquet-files/) work.
all_files: list[str] = []
for dataset_file in local_datasets:
file_path = (
dataset_file
if os.path.isabs(dataset_file)
else os.path.join(
str(datasets_root()),
dataset_file,
)
)
if os.path.isdir(file_path):
file_path_obj = Path(file_path)
parquet_dir = (
file_path_obj / "parquet-files"
if (file_path_obj / "parquet-files").exists()
else file_path_obj
)
parquet_files = sorted(parquet_dir.glob("*.parquet"))
if parquet_files:
all_files.extend(str(p) for p in parquet_files)
continue
candidates: list[Path] = []
for ext in (".json", ".jsonl", ".csv", ".parquet"):
candidates.extend(sorted(file_path_obj.glob(f"*{ext}")))
if candidates:
all_files.extend(str(c) for c in candidates)
continue
raise ValueError(
f"No supported data files in directory: {file_path_obj}"
)
else:
all_files.append(file_path)
if all_files:
first_ext = Path(all_files[0]).suffix.lower()
if first_ext in (".json", ".jsonl"):
loader = "json"
elif first_ext == ".csv":
loader = "csv"
elif first_ext == ".parquet":
loader = "parquet"
else:
raise ValueError(
f"Unsupported local dataset format: {all_files[0]}"
)
dataset = load_dataset(loader, data_files = all_files, split = "train")
else:
event_queue.put(
{
"type": "error",
"error": "No dataset specified for embedding training.",
"stack": "",
"ts": time.time(),
}
)
return
# Apply dataset slicing if specified
slice_start = config.get("dataset_slice_start")
slice_end = config.get("dataset_slice_end")
if slice_start is not None or slice_end is not None:
start = slice_start if slice_start is not None else 0
end = slice_end if slice_end is not None else len(dataset)
dataset = dataset.select(range(start, min(end + 1, len(dataset))))
logger.info(f"Embedding dataset loaded: {len(dataset)} samples")
except Exception as e:
event_queue.put(
{
"type": "error",
"error": f"Failed to load dataset: {e}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
if _should_stop:
event_queue.put({"type": "complete", "output_dir": None, "ts": time.time()})
return
# ── 5. Create loss function ──
loss = MultipleNegativesRankingLoss(model)
# ── 6. Build training arguments ──
_send_status(event_queue, "Configuring training...")
try:
lr_value = float(config.get("learning_rate", "2e-4"))
except ValueError:
event_queue.put(
{
"type": "error",
"error": f"Invalid learning rate: {config.get('learning_rate')}",
"stack": "",
"ts": time.time(),
}
)
return
resume_from_checkpoint = config.get("resume_from_checkpoint")
output_dir = config.get("output_dir") or _output_dir_from_resume_checkpoint(
resume_from_checkpoint
)
if not output_dir:
output_dir = str(
resolve_output_dir(f"{model_name.replace('/', '_')}_{int(time.time())}")
)
output_dir = str(resolve_output_dir(output_dir))
num_epochs = config.get("num_epochs", 2)
batch_size = config.get("batch_size", 256)
gradient_accumulation_steps = config.get("gradient_accumulation_steps", 1)
max_steps_val = config.get("max_steps", 0)
save_steps_val = config.get("save_steps", 0)
warmup_ratio = config.get("warmup_ratio", 0.03)
warmup_steps_val = config.get("warmup_steps")
log_frequency = config.get("log_frequency", 50)
# Build args dict
training_args_kwargs = {
"output_dir": output_dir,
"per_device_train_batch_size": batch_size,
"gradient_accumulation_steps": gradient_accumulation_steps,
"learning_rate": lr_value,
"fp16": not is_bfloat16_supported(),
"bf16": is_bfloat16_supported(),
"logging_steps": 1,
"report_to": ["wandb"] if config.get("enable_wandb") else "none",
"lr_scheduler_type": config.get("lr_scheduler_type", "linear"),
"batch_sampler": BatchSamplers.NO_DUPLICATES,
"optim": config.get("optim", "adamw_8bit"),
"weight_decay": config.get("weight_decay", 0.001),
"seed": config.get("random_seed", 3407),
}
# max_steps vs epochs
if max_steps_val and max_steps_val > 0:
training_args_kwargs["max_steps"] = max_steps_val
else:
training_args_kwargs["num_train_epochs"] = num_epochs if num_epochs > 0 else 2
# warmup: prefer warmup_ratio (standard for embedding scripts), fallback to steps
if warmup_ratio is not None and warmup_ratio > 0:
training_args_kwargs["warmup_ratio"] = warmup_ratio
elif warmup_steps_val is not None and warmup_steps_val > 0:
training_args_kwargs["warmup_steps"] = warmup_steps_val
# save_steps
if save_steps_val and save_steps_val > 0:
training_args_kwargs["save_steps"] = save_steps_val
training_args_kwargs["save_strategy"] = "steps"
args = SentenceTransformerTrainingArguments(**training_args_kwargs)
# ── 7. Calculate total steps for progress tracking ──
if max_steps_val and max_steps_val > 0:
total_steps = max_steps_val
else:
effective_epochs = num_epochs if num_epochs > 0 else 2
len_dataloader = math.ceil(len(dataset) / batch_size)
steps_per_epoch = max(len_dataloader // gradient_accumulation_steps, 1)
total_steps = steps_per_epoch * effective_epochs
# ── 8. Create progress callback ──
class _EmbeddingProgressCallback(TrainerCallback):
"""Sends training progress events to the parent process via event_queue."""
def on_log(self, args, state, control, logs = None, **kwargs):
if not logs:
return
loss_value = logs.get("loss", logs.get("train_loss", None))
current_step = state.global_step
elapsed = time.time() - training_start_time
eta = None
if current_step > 0 and total_steps > 0:
remaining = total_steps - current_step
if remaining > 0:
eta = (elapsed / current_step) * remaining
event_queue.put(
{
"type": "progress",
"step": current_step,
"epoch": round(state.epoch, 2) if state.epoch else 0,
"loss": loss_value,
"learning_rate": logs.get("learning_rate", None),
"total_steps": total_steps,
"elapsed_seconds": elapsed,
"eta_seconds": eta,
"grad_norm": logs.get("grad_norm"),
"num_tokens": getattr(state, "num_input_tokens_seen", None),
"eval_loss": logs.get("eval_loss"),
"status_message": "",
"ts": time.time(),
}
)
def on_step_end(self, args, state, control, **kwargs):
if _should_stop:
logger.info("Embedding training: stop at step %d", state.global_step)
control.should_training_stop = True
return control
# ── 9. Create trainer and train ──
_send_status(event_queue, "Starting embedding training...")
try:
trainer = SentenceTransformerTrainer(
model = model,
train_dataset = dataset,
loss = loss,
args = args,
callbacks = [_EmbeddingProgressCallback()],
)
trainer.train(resume_from_checkpoint = resume_from_checkpoint)
except Exception as e:
event_queue.put(
{
"type": "error",
"error": f"Embedding training failed: {e}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 10. Save model ──
if _should_stop and not _save_on_stop:
event_queue.put(
{
"type": "complete",
"output_dir": None,
"status_message": "Training cancelled",
"ts": time.time(),
}
)
return
_send_status(event_queue, "Saving model...")
try:
if _should_stop and _save_on_stop:
trainer._save_checkpoint(trainer.model, trial = None)
model.save_pretrained(output_dir)
model.tokenizer.save_pretrained(output_dir)
logger.info("Embedding model saved to %s", output_dir)
except Exception as e:
logger.error("Failed to save embedding model: %s", e)
event_queue.put(
{
"type": "error",
"error": f"Training completed but failed to save: {e}",
"stack": traceback.format_exc(limit = 20),
"ts": time.time(),
}
)
return
# ── 11. Done ──
event_queue.put(
{
"type": "complete",
"output_dir": output_dir,
"status_message": "Embedding training completed",
"ts": time.time(),
}
)