[Model][NVIDIA] Route DSA models to the CUDA non-compiled path (#52861)

This commit is contained in:
Woosuk Kwon
2026-08-19 00:13:52 -07:00
committed by GitHub
parent ee11730751
commit b09bd69b5b
16 changed files with 260 additions and 146 deletions
+1 -20
View File
@@ -90,11 +90,8 @@ def run_e2e_fusion_test(monkeypatch, caplog_mp_spawn):
backend_name = attn_backend.backend.name.lower()
requires_mla = "deepseek" in model_name.lower()
is_mla = "mla" in backend_name
# DeepSeek V3.2 uses sparse MLA
requires_sparse = "v3.2" in model_name.lower()
is_sparse = "sparse" in backend_name
if requires_mla != is_mla or requires_sparse != is_sparse:
if requires_mla != is_mla:
pytest.skip(
f"Incompatible model '{model_name}' and "
f"attention backend '{attn_backend.backend.name}'"
@@ -128,22 +125,6 @@ def run_e2e_fusion_test(monkeypatch, caplog_mp_spawn):
# models (e.g. Llama-4-Scout-FP8) at 16384 tokens may trigger OOM.
model_kwargs.setdefault("max_num_batched_tokens", 8192)
# Sparse MLA models (DSv3.2) hit an over-strict inductor assertion in
# decompose_auto_functionalized when +rotary_embedding is forced into
# the compile graph. Disable qk_norm+rope fusion (which auto-enables
# +rotary_embedding) for this combo to avoid the known torch bug.
# TODO: remove once upstream torch fix lands.
if requires_sparse:
if "pass_config" in compilation_config:
compilation_config["pass_config"].enable_qk_norm_rope_fusion = False
matches_check = [m for m in matches_check if m != "norm_rope_fusion"]
# DSv3.2 sparse indexer uses persistent_topk with k=config.index_topk
# (2048 for the default config). max_model_len must be >= index_topk
# or the topk kernel raises "k out of range" at runtime.
model_kwargs["max_model_len"] = max(
model_kwargs.get("max_model_len", 0), 2048
)
# Always compile the full graph instead of piecewise
if not compilation_config["use_inductor_graph_partition"]:
compilation_config["splitting_ops"] = []