[ROCm][DSv4.1] Disambiguate ROCm attention class names (#58689)

Signed-off-by: Tuukka Sarvi <tuukka.sarvi@amd.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Tuukka Sarvi
2026-10-15 21:38:39 +08:00
committed by GitHub
co-authored by Shanshan Shen Kevin H. Luu
parent 1463501553
commit 0951fc183c
11 changed files with 86 additions and 29 deletions
@@ -1989,8 +1989,8 @@ def test_dsv41_adaptive_mla_swa_metadata_graph_replay(monkeypatch) -> None:
"""Replay V4.1 ROCm decode after device-only query reallocation."""
from tests.v1.attention.utils import create_vllm_config
from vllm.models.deepseek_v41.amd.rocm import (
DeepseekV4ROCMAiterSparseSWAMetadataBuilder,
DeepseekV41ROCMAiterMLASparseMetadataBuilder,
DeepseekV41ROCMAiterSparseSWAMetadataBuilder,
compute_global_topk_ragged_indices_and_indptr,
)
from vllm.v1.attention.backend import AttentionCGSupport, CommonAttentionMetadata
@@ -2044,7 +2044,7 @@ def test_dsv41_adaptive_mla_swa_metadata_graph_replay(monkeypatch) -> None:
mla_builder = DeepseekV41ROCMAiterMLASparseMetadataBuilder(
mla_spec, ["c2a"], vllm_config, device
)
swa_builder = DeepseekV4ROCMAiterSparseSWAMetadataBuilder(
swa_builder = DeepseekV41ROCMAiterSparseSWAMetadataBuilder(
swa_spec, ["c2a"], vllm_config, device
)
assert (
@@ -255,7 +255,7 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
vllm_runner_kwargs={
"load_format": "dummy",
"attention_backend": (
"ROCM_FLASHMLA_SPARSE_DSV4"
"ROCM_FLASHMLA_SPARSE_DSV41"
if current_platform.is_rocm()
else "FLASHMLA_SPARSE_DSV41"
),
@@ -67,7 +67,7 @@ def test_deepseek_rocm_adaptive_builders_support_varlen_full_graphs():
DeepseekV4ROCMAiterMLASparseMetadataBuilder,
DeepseekV4ROCMAiterSparseSWAMetadataBuilder,
DeepseekV41ROCMAiterMLASparseMetadataBuilder,
v41_rocm.DeepseekV4ROCMAiterSparseSWAMetadataBuilder,
v41_rocm.DeepseekV41ROCMAiterSparseSWAMetadataBuilder,
):
assert (
builder_cls.get_cudagraph_support(adaptive_config, SimpleNamespace())
@@ -126,7 +126,7 @@ def test_deepseek_v41_rocm_fixed_k_path_is_unchanged(
for builder_cls in (
DeepseekV41ROCMAiterMLASparseMetadataBuilder,
v41_rocm.DeepseekV4ROCMAiterSparseSWAMetadataBuilder,
v41_rocm.DeepseekV41ROCMAiterSparseSWAMetadataBuilder,
):
assert (
builder_cls.get_cudagraph_support(fixed_config, SimpleNamespace())
@@ -171,6 +171,11 @@ def test_paged_prefill_indices_match_gathered_prefill(
from vllm.v1.kv_cache_interface import SlidingWindowMLASpec
rocm = importlib.import_module(f"vllm.models.{model}.amd.rocm")
builder_cls = (
rocm.DeepseekV41ROCMAiterSparseSWAMetadataBuilder
if model == "deepseek_v41"
else rocm.DeepseekV4ROCMAiterSparseSWAMetadataBuilder
)
device = torch.device("cuda")
gen = torch.Generator().manual_seed(compress_ratio)
window, width, block_size, max_model_len = 16, 32, 16, 128
@@ -233,9 +238,7 @@ def test_paged_prefill_indices_match_gathered_prefill(
monkeypatch.setattr(
rocm_aiter_ops, "is_triton_sparse_mla_enabled", lambda: enabled
)
return rocm.DeepseekV4ROCMAiterSparseSWAMetadataBuilder(
swa_spec, ["swa"], vllm_config, device
).build(0, common)
return builder_cls(swa_spec, ["swa"], vllm_config, device).build(0, common)
assert build(enabled=False).prefill_swa_ragged_indices is None
md = build(enabled=True)
@@ -8,6 +8,7 @@ from unittest.mock import MagicMock, patch
import pytest
import torch
from vllm.config.attention import AttentionConfig
from vllm.platforms import current_platform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.selector import AttentionSelectorConfig
@@ -44,6 +45,47 @@ def cleared_attention_selector_cache():
_cached_get_attn_backend.cache_clear()
def test_deepseek_rocm_backend_registry_paths_are_model_specific():
assert AttentionBackendEnum.ROCM_FLASHMLA_SPARSE_DSV4.get_path() == (
"vllm.models.deepseek_v4.amd.rocm.DeepseekV4ROCMAiterMLASparseBackend"
)
assert AttentionBackendEnum.ROCM_FLASHMLA_SPARSE_DSV41.get_path() == (
"vllm.models.deepseek_v41.amd.rocm.DeepseekV41ROCMAiterMLASparseBackend"
)
@pytest.mark.parametrize(
"backend_name, expect_warning",
[
("ROCM_FLASHMLA_SPARSE_DSV4", True),
("ROCM_FLASHMLA_SPARSE_DSV41", False),
],
ids=["legacy", "canonical"],
)
def test_deepseek_v41_rocm_accepts_legacy_backend_alias(backend_name, expect_warning):
from vllm.models.deepseek_v41.amd import model as dsv41_model
from vllm.models.deepseek_v41.amd.rocm import (
DeepseekV41ROCMAiterMLAAttention,
DeepseekV41ROCMAiterMLASparseBackend,
)
config = SimpleNamespace(
attention_config=AttentionConfig(backend=backend_name),
)
with patch.object(dsv41_model.logger, "warning_once") as warning_once:
selected = dsv41_model._select_dsv4_attn_cls(config)
assert selected is DeepseekV41ROCMAiterMLAAttention
if expect_warning:
warning_once.assert_called_once()
else:
warning_once.assert_not_called()
assert (
DeepseekV41ROCMAiterMLASparseBackend.get_name() == "ROCM_FLASHMLA_SPARSE_DSV41"
)
def test_aiter_unified_attention_uses_dedicated_metadata_builder():
from vllm.v1.attention.backends.rocm_aiter_unified_attn import (
RocmAiterUnifiedAttentionBackend,