[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
@@ -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 = []