Add RDNA 2/3/4 ROCm routing tests via a CPU-only torch spoof (#6935)
* Add RDNA 2/3/4 ROCm routing tests via a CPU-only torch spoof Introduces tests/_zoo_rocm_spoof.py, the ROCm sibling of _zoo_aggressive_cuda_spoof.py: it reuses the CUDA spoof's torch.cuda no-op machinery and overlays an AMD Radeon identity (torch.version.hip, gcnArchName, capability) for any RDNA 2/3/4 gfx target, so hip code paths run on CPU-only CI with no AMD hardware. tests/studio/install/test_rocm_rdna_routing.py then asserts unsloth_zoo routes every RDNA arch (gfx1030/1031/1032/1034, gfx1100/1101/1102, gfx1150/1151, gfx1200/1201) correctly: device_type resolves to hip, llama.cpp target resolves to (rocm, gfx), and the per-family ROCm bundle suffix (gfx103X/gfx110X/gfx120X, or self for gfx1150/1151) is picked. The torch-facing checks run in a subprocess so the spoof never leaks into sibling tests and DEVICE_TYPE (cached at import) resolves from a clean process; the pure gfx-family mapping runs in-process. Guarded by importorskip so it runs where torch and unsloth_zoo are installed (the Repo tests CPU job) and skips elsewhere. * [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>
This commit is contained in:
Daniel Han
GitHub
parent
414503745e
commit
d79495dc96
2 changed files with 168 additions and 0 deletions
84
tests/_zoo_rocm_spoof.py
Normal file
84
tests/_zoo_rocm_spoof.py
Normal file
|
|
@ -0,0 +1,84 @@
|
|||
# SPDX-License-Identifier: AGPL-3.0-only
|
||||
# Copyright 2026-present the Unsloth AI Inc. team.
|
||||
"""ROCm/RDNA spoof: present torch as an AMD Radeon (RDNA 2/3/4) card on a
|
||||
GPU-less host, so hip paths (device_type -> "hip", llama.cpp ROCm bundle) are
|
||||
testable in CPU-only CI with no AMD hardware. The ROCm sibling of
|
||||
_zoo_aggressive_cuda_spoof.py: it reuses that spoof's torch.cuda no-op machinery
|
||||
and overlays the AMD identity (torch.version.hip, gcnArchName, Radeon name).
|
||||
Apply BEFORE importing unsloth/unsloth_zoo, since DEVICE_TYPE is cached there.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import importlib.util
|
||||
import os
|
||||
import sys
|
||||
|
||||
# gfx -> (marketing name, (capability major, minor), torch.version.hip). hip is
|
||||
# the ROCm build torch was made against (RDNA2/3 ship 6.x; gfx1102/115x/RDNA4 7.2).
|
||||
_PROFILES: dict[str, tuple[str, tuple[int, int], str]] = {
|
||||
"gfx1030": ("AMD Radeon RX 6900 XT", (10, 3), "6.4.43483"), # RDNA2
|
||||
"gfx1031": ("AMD Radeon RX 6700 XT", (10, 3), "6.4.43483"),
|
||||
"gfx1032": ("AMD Radeon RX 6600", (10, 3), "6.4.43483"),
|
||||
"gfx1034": ("AMD Radeon RX 6400", (10, 3), "6.4.43483"),
|
||||
"gfx1100": ("AMD Radeon RX 7900 XTX", (11, 0), "6.4.43483"), # RDNA3
|
||||
"gfx1101": ("AMD Radeon RX 7800 XT", (11, 0), "6.4.43483"),
|
||||
"gfx1102": ("AMD Radeon RX 7600", (11, 0), "7.2.1"),
|
||||
"gfx1150": ("AMD Radeon 890M", (11, 5), "7.2.1"), # RDNA3.5 APU
|
||||
"gfx1151": ("AMD Radeon 8060S", (11, 5), "7.2.1"),
|
||||
"gfx1200": ("AMD Radeon RX 9060 XT", (12, 0), "7.2.1"), # RDNA4
|
||||
"gfx1201": ("AMD Radeon RX 9070 XT", (12, 0), "7.2.1"),
|
||||
}
|
||||
|
||||
|
||||
def _cuda_spoof():
|
||||
"""Load the sibling CUDA spoof by path (robust to sys.path), so we reuse its
|
||||
torch.cuda machinery instead of duplicating it."""
|
||||
if "_zoo_aggressive_cuda_spoof" in sys.modules:
|
||||
return sys.modules["_zoo_aggressive_cuda_spoof"]
|
||||
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "_zoo_aggressive_cuda_spoof.py")
|
||||
spec = importlib.util.spec_from_file_location("_zoo_aggressive_cuda_spoof", path)
|
||||
mod = importlib.util.module_from_spec(spec)
|
||||
spec.loader.exec_module(mod)
|
||||
sys.modules["_zoo_aggressive_cuda_spoof"] = mod
|
||||
return mod
|
||||
|
||||
|
||||
def apply(gfx: str = "gfx1100", device_count: int = 1) -> None:
|
||||
"""Present torch as `gfx`. Re-callable to switch arch (identity is overlaid;
|
||||
the underlying no-op machinery is applied once)."""
|
||||
import torch
|
||||
|
||||
if gfx not in _PROFILES:
|
||||
raise KeyError(f"Unknown gfx {gfx!r}; known: {', '.join(_PROFILES)}")
|
||||
name, cap, hip = _PROFILES[gfx]
|
||||
|
||||
_cuda_spoof().apply() # is_available/device_count/streams/rng/amp/...
|
||||
|
||||
# Overlay the AMD identity on top of the (NVIDIA-shaped) CUDA spoof.
|
||||
torch.version.hip = hip
|
||||
torch.version.cuda = None
|
||||
torch.cuda.device_count = lambda: device_count
|
||||
torch.cuda.get_device_name = lambda *a, **k: name
|
||||
torch.cuda.get_device_capability = lambda *a, **k: cap
|
||||
torch.cuda.get_arch_list = lambda: [gfx]
|
||||
|
||||
class _Props:
|
||||
pass
|
||||
|
||||
_p = _Props()
|
||||
_p.name = name
|
||||
_p.gcnArchName = f"{gfx}:sramecc-:xnack-" # ROCm advertises feature flags
|
||||
_p.major, _p.minor = cap
|
||||
_p.total_memory = 16 * 1024**3
|
||||
_p.multi_processor_count = 40
|
||||
_p.warp_size = 32 # RDNA wavefront (CDNA is 64)
|
||||
_p.is_integrated = gfx in ("gfx1150", "gfx1151")
|
||||
_p.is_multi_gpu_board = False
|
||||
torch.cuda.get_device_properties = lambda *a, **k: _p
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
apply()
|
||||
import torch
|
||||
print("ROCm spoof applied:", torch.version.hip, torch.cuda.get_device_properties(0).gcnArchName)
|
||||
84
tests/studio/install/test_rocm_rdna_routing.py
Normal file
84
tests/studio/install/test_rocm_rdna_routing.py
Normal file
|
|
@ -0,0 +1,84 @@
|
|||
# SPDX-License-Identifier: AGPL-3.0-only
|
||||
# Copyright 2026-present the Unsloth AI Inc. team.
|
||||
"""RDNA 2/3/4 routing, validated on CPU-only CI with no AMD hardware.
|
||||
|
||||
tests/_zoo_rocm_spoof.py presents torch as each Radeon gfx arch, then we assert
|
||||
unsloth_zoo routes it: device_type -> "hip", llama.cpp target -> ("rocm", gfx),
|
||||
and the per-family ROCm bundle suffix. The torch-facing checks run in a
|
||||
subprocess so the spoof never leaks into sibling tests and DEVICE_TYPE (cached
|
||||
at import) resolves from a clean process.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import subprocess
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
|
||||
pytest.importorskip("torch")
|
||||
pytest.importorskip("unsloth_zoo")
|
||||
|
||||
_TESTS_DIR = Path(__file__).resolve().parents[2] # tests/
|
||||
|
||||
# gfx -> (expected llama.cpp target, expected ROCm bundle family).
|
||||
_ARCHES = {
|
||||
"gfx1030": (("rocm", "gfx1030"), "gfx103X"), # RDNA2
|
||||
"gfx1031": (("rocm", "gfx1031"), "gfx103X"),
|
||||
"gfx1032": (("rocm", "gfx1032"), "gfx103X"),
|
||||
"gfx1034": (("rocm", "gfx1034"), "gfx103X"),
|
||||
"gfx1100": (("rocm", "gfx1100"), "gfx110X"), # RDNA3
|
||||
"gfx1101": (("rocm", "gfx1101"), "gfx110X"),
|
||||
"gfx1102": (("rocm", "gfx1102"), "gfx110X"),
|
||||
"gfx1150": (("rocm", "gfx1150"), "gfx1150"), # RDNA3.5 APU (self-family)
|
||||
"gfx1151": (("rocm", "gfx1151"), "gfx1151"),
|
||||
"gfx1200": (("rocm", "gfx1200"), "gfx120X"), # RDNA4
|
||||
"gfx1201": (("rocm", "gfx1201"), "gfx120X"),
|
||||
}
|
||||
|
||||
# Child: spoof each arch, then record device_type once (fresh import) and the
|
||||
# live llama.cpp target per arch. Emits one JSON line the parent parses.
|
||||
_CHILD = """
|
||||
import json, sys
|
||||
sys.path.insert(0, {tests!r})
|
||||
import _zoo_rocm_spoof as spoof
|
||||
arches = {arches!r}
|
||||
spoof.apply(arches[0])
|
||||
from unsloth_zoo.device_type import get_device_type, is_hip
|
||||
device_type = [get_device_type(), is_hip()]
|
||||
from unsloth_zoo import llama_cpp as lc
|
||||
targets = {{}}
|
||||
for gfx in arches:
|
||||
spoof.apply(gfx)
|
||||
targets[gfx] = list(lc._detect_gpu_target())
|
||||
print("RESULT " + json.dumps({{"device_type": device_type, "targets": targets}}))
|
||||
"""
|
||||
|
||||
|
||||
@pytest.fixture(scope = "module")
|
||||
def routed():
|
||||
code = _CHILD.format(tests = str(_TESTS_DIR), arches = list(_ARCHES))
|
||||
proc = subprocess.run([sys.executable, "-c", code], capture_output = True, text = True)
|
||||
line = next((l for l in proc.stdout.splitlines() if l.startswith("RESULT ")), None)
|
||||
assert line, f"child produced no result.\nstdout:\n{proc.stdout}\nstderr:\n{proc.stderr}"
|
||||
return json.loads(line[len("RESULT ") :])
|
||||
|
||||
|
||||
@pytest.mark.parametrize("gfx", list(_ARCHES))
|
||||
def test_detect_gpu_target(routed, gfx):
|
||||
# RDNA card is routed to its ROCm gfx target (drives the llama.cpp bundle).
|
||||
assert tuple(routed["targets"][gfx]) == _ARCHES[gfx][0]
|
||||
|
||||
|
||||
def test_device_type_is_hip(routed):
|
||||
# An RDNA card must resolve the compute device_type to "hip".
|
||||
assert routed["device_type"] == ["hip", True]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("gfx", list(_ARCHES))
|
||||
def test_rocm_gfx_family(gfx):
|
||||
# Pure mapping (no torch): each gfx picks the right per-family ROCm bundle.
|
||||
from unsloth_zoo import llama_cpp as lc
|
||||
assert lc._rocm_gfx_family(gfx) == _ARCHES[gfx][1]
|
||||
Loading…
Add table
Add a link
Reference in a new issue