[CI] fix compile test | refactor VLLM_DISABLE_COMPILE_CACHE for tests (#49770)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
This commit is contained in:
@@ -290,12 +290,9 @@ def test_attention_quant_pattern(
|
|||||||
model_class: type[AttentionQuantPatternModel],
|
model_class: type[AttentionQuantPatternModel],
|
||||||
backend: AttentionBackendEnum,
|
backend: AttentionBackendEnum,
|
||||||
dist_init,
|
dist_init,
|
||||||
monkeypatch,
|
disable_vllm_compile_cache,
|
||||||
use_fresh_inductor_cache,
|
|
||||||
):
|
):
|
||||||
"""Test AttentionStaticQuantPattern fusion pass"""
|
"""Test AttentionStaticQuantPattern fusion pass"""
|
||||||
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
|
|
||||||
|
|
||||||
if backend == AttentionBackendEnum.FLASHINFER and (
|
if backend == AttentionBackendEnum.FLASHINFER and (
|
||||||
not current_platform.is_device_capability((10, 0)) or not has_flashinfer()
|
not current_platform.is_device_capability((10, 0)) or not has_flashinfer()
|
||||||
):
|
):
|
||||||
|
|||||||
@@ -419,8 +419,7 @@ def test_mla_attention_quant_pattern(
|
|||||||
model_class: type[MLAAttentionQuantPatternModel],
|
model_class: type[MLAAttentionQuantPatternModel],
|
||||||
backend: AttentionBackendEnum,
|
backend: AttentionBackendEnum,
|
||||||
dist_init,
|
dist_init,
|
||||||
monkeypatch,
|
disable_vllm_compile_cache,
|
||||||
use_fresh_inductor_cache,
|
|
||||||
):
|
):
|
||||||
"""Test MLA AttentionQuantPattern fusion pass"""
|
"""Test MLA AttentionQuantPattern fusion pass"""
|
||||||
if (
|
if (
|
||||||
@@ -429,8 +428,6 @@ def test_mla_attention_quant_pattern(
|
|||||||
):
|
):
|
||||||
pytest.skip("NVFP4 is not supported on this GPU (requires SM 100+).")
|
pytest.skip("NVFP4 is not supported on this GPU (requires SM 100+).")
|
||||||
|
|
||||||
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
|
|
||||||
|
|
||||||
custom_ops_list = custom_ops.split(",") if custom_ops else []
|
custom_ops_list = custom_ops.split(",") if custom_ops else []
|
||||||
|
|
||||||
device = torch.device(f"{DEVICE_TYPE}:0")
|
device = torch.device(f"{DEVICE_TYPE}:0")
|
||||||
|
|||||||
@@ -66,7 +66,7 @@ class PostGradRangeChecker(InductorPass):
|
|||||||
return InductorPass.hash_dict(state)
|
return InductorPass.hash_dict(state)
|
||||||
|
|
||||||
|
|
||||||
def test_compile_ranges(use_fresh_inductor_cache):
|
def test_compile_ranges(disable_vllm_compile_cache):
|
||||||
post_grad_range_checker = PostGradRangeChecker(
|
post_grad_range_checker = PostGradRangeChecker(
|
||||||
[
|
[
|
||||||
Range(start=1, end=8),
|
Range(start=1, end=8),
|
||||||
@@ -168,7 +168,7 @@ class PostGradStaticShapeChecker(InductorPass):
|
|||||||
return InductorPass.hash_dict(state)
|
return InductorPass.hash_dict(state)
|
||||||
|
|
||||||
|
|
||||||
def test_compile_sizes_produce_static_shapes(use_fresh_inductor_cache):
|
def test_compile_sizes_produce_static_shapes(disable_vllm_compile_cache):
|
||||||
"""Verify that compile_sizes entries are compiled with fully concrete
|
"""Verify that compile_sizes entries are compiled with fully concrete
|
||||||
shapes (no SymInts), while compile_ranges entries retain dynamic shapes."""
|
shapes (no SymInts), while compile_ranges entries retain dynamic shapes."""
|
||||||
checker = PostGradStaticShapeChecker()
|
checker = PostGradStaticShapeChecker()
|
||||||
@@ -209,10 +209,9 @@ def test_compile_sizes_produce_static_shapes(use_fresh_inductor_cache):
|
|||||||
)
|
)
|
||||||
|
|
||||||
|
|
||||||
def test_inductor_cache_compile_ranges(monkeypatch, use_fresh_inductor_cache):
|
def test_inductor_cache_compile_ranges(disable_vllm_compile_cache):
|
||||||
# To force multiple compilations, we disable the compile cache
|
# disable_vllm_compile_cache sets VLLM_DISABLE_COMPILE_CACHE=1 to force
|
||||||
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
|
# multiple compilations by disabling vLLM's on-disk compile cache.
|
||||||
|
|
||||||
post_grad_range_checker = PostGradRangeChecker(
|
post_grad_range_checker = PostGradRangeChecker(
|
||||||
ranges=[
|
ranges=[
|
||||||
Range(start=1, end=8),
|
Range(start=1, end=8),
|
||||||
|
|||||||
@@ -1768,6 +1768,22 @@ def use_fresh_inductor_cache():
|
|||||||
yield
|
yield
|
||||||
|
|
||||||
|
|
||||||
|
@pytest.fixture
|
||||||
|
def disable_vllm_compile_cache(monkeypatch, use_fresh_inductor_cache):
|
||||||
|
"""
|
||||||
|
Use a fresh inductor cache AND disable vLLM's on-disk torch.compile cache.
|
||||||
|
|
||||||
|
This forces compilation (and any custom compile passes) to actually run
|
||||||
|
instead of being served from a warm cache left behind by previous runs
|
||||||
|
(e.g. on persistent CI agents). Use this for tests that inspect what
|
||||||
|
happens during compilation; use ``use_fresh_inductor_cache`` (or
|
||||||
|
``fresh_vllm_cache``) instead when the vLLM compile cache must stay
|
||||||
|
enabled (e.g. cache save/load tests).
|
||||||
|
"""
|
||||||
|
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
|
||||||
|
yield
|
||||||
|
|
||||||
|
|
||||||
@pytest.fixture
|
@pytest.fixture
|
||||||
def fresh_vllm_cache(monkeypatch, use_fresh_inductor_cache):
|
def fresh_vllm_cache(monkeypatch, use_fresh_inductor_cache):
|
||||||
"""Temporary VLLM_CACHE_ROOT combined with a fresh inductor cache."""
|
"""Temporary VLLM_CACHE_ROOT combined with a fresh inductor cache."""
|
||||||
|
|||||||
Reference in New Issue
Block a user