[ROCm][Bugfix] Reduce CUDA graph divergences (#57229)
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
This commit is contained in:
@@ -477,45 +477,6 @@ def test_rocm_keeps_compiled_deepseek_defaults(monkeypatch):
|
||||
default_breakable_cudagraph_architectures.cache_clear()
|
||||
|
||||
|
||||
@pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm-specific test")
|
||||
@pytest.mark.parametrize(
|
||||
("architecture", "use_v2", "mode", "expected"),
|
||||
[
|
||||
("DeepseekV4ForCausalLM", True, None, CUDAGraphMode.NONE),
|
||||
("DeepseekV4ForConditionalGeneration", True, None, CUDAGraphMode.NONE),
|
||||
("DeepseekV4ForCausalLM", False, None, None),
|
||||
("LlamaForCausalLM", True, None, None),
|
||||
(
|
||||
"DeepseekV4ForCausalLM",
|
||||
True,
|
||||
CUDAGraphMode.FULL_DECODE_ONLY,
|
||||
CUDAGraphMode.FULL_DECODE_ONLY,
|
||||
),
|
||||
],
|
||||
)
|
||||
def test_rocm_gfx950_deepseek_v4_cudagraph_default(
|
||||
monkeypatch, architecture, use_v2, mode, expected
|
||||
):
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.platforms import rocm
|
||||
|
||||
monkeypatch.setattr(rocm, "on_gfx950", lambda: True)
|
||||
monkeypatch.setattr(rocm_aiter_ops, "is_fused_moe_enabled", lambda: False)
|
||||
monkeypatch.setattr(rocm_aiter_ops, "is_linear_fp8_enabled", lambda: False)
|
||||
monkeypatch.setattr(
|
||||
rocm_aiter_ops, "is_fusion_moe_shared_experts_enabled", lambda: False
|
||||
)
|
||||
config = SimpleNamespace(
|
||||
compilation_config=CompilationConfig(cudagraph_mode=mode),
|
||||
model_config=SimpleNamespace(architecture=architecture),
|
||||
use_v2_model_runner=use_v2,
|
||||
)
|
||||
|
||||
rocm.RocmPlatform.apply_config_platform_defaults(config)
|
||||
|
||||
assert config.compilation_config.cudagraph_mode == expected
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("model", "architecture"),
|
||||
[
|
||||
|
||||
Reference in New Issue
Block a user