[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
-12
View File
@@ -50,7 +50,6 @@ MODELS = ["TinyLlama/TinyLlama-1.1B-Chat-v1.0"]
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
@pytest.mark.parametrize("beam_width", BEAM_WIDTHS)
def test_beam_search_single_input(
monkeypatch,
hf_runner,
vllm_runner,
example_prompts,
@@ -59,9 +58,6 @@ def test_beam_search_single_input(
max_tokens: int,
beam_width: int,
) -> None:
if current_platform.is_rocm():
monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "0")
example_prompts = example_prompts[:1]
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_beam_search(
@@ -95,7 +91,6 @@ def test_beam_search_single_input(
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
@pytest.mark.parametrize("beam_width", BEAM_WIDTHS)
def test_beam_search_with_concurrency_limit(
monkeypatch,
hf_runner,
vllm_runner,
example_prompts,
@@ -104,9 +99,6 @@ def test_beam_search_with_concurrency_limit(
max_tokens: int,
beam_width: int,
) -> None:
if current_platform.is_rocm():
monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "0")
# example_prompts[1]&[3]&[7] fails due to unknown reason even without
# concurrency limit. skip them for now.
example_prompts = example_prompts[:8]
@@ -156,7 +148,6 @@ def test_beam_search_with_concurrency_limit(
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
@pytest.mark.parametrize("beam_width", MM_BEAM_WIDTHS)
def test_beam_search_passes_multimodal_data(
monkeypatch,
hf_runner,
vllm_runner,
dtype: str,
@@ -164,9 +155,6 @@ def test_beam_search_passes_multimodal_data(
beam_width: int,
) -> None:
"""Ensure that beam search passes multimodal data through correctly."""
if current_platform.is_rocm():
monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "0")
# NOTE - this test is primarily to check that mm data is passed to beams
# correctly. As such, we just need to check one extra modality to make
# sure things pass through properly.