[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
@@ -10,15 +10,13 @@ import pybase64 as base64
import pytest
import websockets
from tests.entrypoints.speech_to_text.conftest import add_attention_backend
from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
from vllm.assets.audio import AudioAsset
from vllm.multimodal.media.audio import load_audio
# Increase engine iteration timeout for ROCm where first-use JIT compilation
# can exceed the default 60s, causing a silent deadlock in feed_tokens.
REALTIME_ENV_OVERRIDES = {
**ROCM_ENV_OVERRIDES,
"VLLM_ENGINE_ITERATION_TIMEOUT_S": "600",
}
@@ -72,17 +70,13 @@ def mary_had_lamb_audio_chunks() -> list[str]:
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
async def test_multi_chunk_streaming(
model_name, mary_had_lamb_audio_chunks, rocm_aiter_fa_attention
):
async def test_multi_chunk_streaming(model_name, mary_had_lamb_audio_chunks):
"""Test streaming multiple audio chunks before committing."""
server_args = ["--enforce-eager", "--max-model-len", "2048"]
if model_name.startswith("mistralai"):
server_args += MISTRAL_FORMAT_ARGS
add_attention_backend(server_args, rocm_aiter_fa_attention)
with RemoteOpenAIServer(
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
) as remote_server:
@@ -174,7 +168,7 @@ async def test_multi_chunk_streaming(
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
async def test_empty_commit_does_not_crash_engine(
model_name, mary_had_lamb_audio_chunks, rocm_aiter_fa_attention
model_name, mary_had_lamb_audio_chunks
):
"""Test that committing without audio does not crash the engine.
@@ -189,8 +183,6 @@ async def test_empty_commit_does_not_crash_engine(
if model_name.startswith("mistralai"):
server_args += MISTRAL_FORMAT_ARGS
add_attention_backend(server_args, rocm_aiter_fa_attention)
with RemoteOpenAIServer(
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
) as remote_server:
@@ -273,17 +265,13 @@ async def test_empty_commit_does_not_crash_engine(
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
async def test_session_update_invalid_model_returns_error(
model_name, rocm_aiter_fa_attention
):
async def test_session_update_invalid_model_returns_error(model_name):
"""Test that session.update with an invalid model returns an error."""
server_args = ["--enforce-eager", "--max-model-len", "2048"]
if model_name.startswith("mistralai"):
server_args += MISTRAL_FORMAT_ARGS
add_attention_backend(server_args, rocm_aiter_fa_attention)
with RemoteOpenAIServer(
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
) as remote_server:
@@ -305,9 +293,7 @@ async def test_session_update_invalid_model_returns_error(
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
async def test_commit_without_session_update_returns_error(
model_name, rocm_aiter_fa_attention
):
async def test_commit_without_session_update_returns_error(model_name):
"""Test that committing before validating the model returns an error
and does not fall through to processing."""
server_args = ["--enforce-eager", "--max-model-len", "2048"]
@@ -315,8 +301,6 @@ async def test_commit_without_session_update_returns_error(
if model_name.startswith("mistralai"):
server_args += MISTRAL_FORMAT_ARGS
add_attention_backend(server_args, rocm_aiter_fa_attention)
with RemoteOpenAIServer(
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
) as remote_server: