[ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
2026-08-11 14:19:16 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f2f31b82e
commit 1ab2801dde
56 changed files with 500 additions and 558 deletions
@@ -1,46 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM language generation tests."""
import os
import warnings
import torch
from vllm.platforms import current_platform
def pytest_configure(config):
"""Early ROCm configuration that must happen before test collection."""
if not current_platform.is_rocm():
return
# Disable skinny GEMM on ROCm to avoid non-deterministic results
# from atomic reductions in wvSplitKrc kernel.
# See: https://github.com/vllm-project/vllm/pull/33493#issuecomment-3906083975
os.environ["VLLM_ROCM_USE_SKINNY_GEMM"] = "0"
warnings.warn(
"ROCm: Set VLLM_ROCM_USE_SKINNY_GEMM=0 to avoid non-deterministic "
"results from skinny GEMM atomic reductions",
UserWarning,
stacklevel=1,
)
def pytest_sessionstart(session):
"""Configure ROCm-specific settings before test session starts."""
if not current_platform.is_rocm():
return
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
# TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_math_sdp(True)
warnings.warn(
"ROCm: Disabled flash_sdp and mem_efficient_sdp, enabled math_sdp "
"to avoid HuggingFace Transformers accuracy issues",
UserWarning,
stacklevel=1,
)
@@ -126,6 +126,12 @@ def test_models(
model_info.check_available_online(on_fail="skip")
model_info.check_transformers_version(on_fail="skip")
if current_platform.is_rocm() and model == "TitanML/tiny-mixtral":
# Its single-token router selects LLMM1, whose low-precision
# accumulation can change the top-2 experts. Keep the optimized kernel
# enabled generally, but use the reference GEMM for this accuracy test.
monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "0")
if use_rocm_aiter and (model in AITER_MODEL_LIST):
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
elif use_rocm_aiter and model not in AITER_MODEL_LIST:
+2 -18
View File
@@ -2,8 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM language generation tests."""
import warnings
import torch
from vllm.platforms import current_platform
@@ -11,19 +9,5 @@ from vllm.platforms import current_platform
def pytest_sessionstart(session):
"""Configure ROCm-specific settings before test session starts."""
if not current_platform.is_rocm():
return
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
# TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_math_sdp(True)
torch.set_float32_matmul_precision("high")
warnings.warn(
"ROCm: Disabled flash_sdp and mem_efficient_sdp, enabled math_sdp "
"to avoid HuggingFace Transformers accuracy issues",
UserWarning,
stacklevel=1,
)
if current_platform.is_rocm():
torch.set_float32_matmul_precision("high")
+2 -2
View File
@@ -7,7 +7,7 @@ from scipy.spatial.distance import cosine
from vllm import LLM, SamplingParams
from vllm.config import ModelConfig
from ....utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
from ....utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
from .embed_utils import run_client_embeddings
MODEL_NAME = "parasail-ai/GritLM-7B-vllm"
@@ -134,7 +134,7 @@ async def test_gritlm_api_server_embedding():
*ROCM_EXTRA_ARGS,
]
with RemoteOpenAIServer(MODEL_NAME, args, env_dict=ROCM_ENV_OVERRIDES) as server:
with RemoteOpenAIServer(MODEL_NAME, args) as server:
client_embedding = server.get_async_client()
d_rep = await run_client_embeddings(
@@ -48,17 +48,10 @@ def test_bert_model_runner_v2(
vllm_model.token_classify(prompts) for prompts in prompt_batches
]
# Use eager attention on ROCm to avoid HF Transformers flash attention
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
hf_model_kwargs = {}
if current_platform.is_rocm():
hf_model_kwargs["attn_implementation"] = "eager"
with hf_runner(
model,
dtype=dtype,
auto_cls=AutoModelForTokenClassification,
model_kwargs=hf_model_kwargs,
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []
@@ -100,17 +93,10 @@ def test_modernbert_models(
with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.token_classify(example_prompts)
# Use eager attention on ROCm to avoid HF Transformers flash attention
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
hf_model_kwargs = {}
if current_platform.is_rocm():
hf_model_kwargs["attn_implementation"] = "eager"
with hf_runner(
model,
dtype=dtype,
auto_cls=AutoModelForTokenClassification,
model_kwargs=hf_model_kwargs,
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []
@@ -140,17 +126,10 @@ def test_xlm_roberta_models(
with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.token_classify(example_prompts)
# Use eager attention on ROCm to avoid HF Transformers flash attention
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
hf_model_kwargs = {}
if current_platform.is_rocm():
hf_model_kwargs["attn_implementation"] = "eager"
with hf_runner(
model,
dtype=dtype,
auto_cls=AutoModelForTokenClassification,
model_kwargs=hf_model_kwargs,
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []
@@ -197,15 +176,10 @@ def test_openai_privacy_filter(
with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.token_classify(PRIVACY_FILTER_PROMPTS)
hf_model_kwargs = {}
if current_platform.is_rocm():
hf_model_kwargs["attn_implementation"] = "eager"
with hf_runner(
model,
dtype=dtype,
auto_cls=AutoModelForTokenClassification,
model_kwargs=hf_model_kwargs,
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []
@@ -276,12 +250,6 @@ def test_bert_for_masked_lm(
with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.token_classify(example_prompts)
# Use eager attention on ROCm to avoid HF Transformers flash attention
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
hf_model_kwargs = {}
if current_platform.is_rocm():
hf_model_kwargs["attn_implementation"] = "eager"
# Run hf_runner reference with "highest" fp32 precision to match
# default behvior of vLLM. This is needed on ROCm since the
# pooling tests set matmul precision to "high" in conftest.py
@@ -291,7 +259,6 @@ def test_bert_for_masked_lm(
model,
dtype=dtype,
auto_cls=AutoModelForMaskedLM,
model_kwargs=hf_model_kwargs,
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []