[BugFix] Rename attention_config to diffusion_attention_config (#3489)

Signed-off-by: gcanlin <canlinguosdu@gmail.com>
This commit is contained in:
Canlin Guo authored and GitHub committed 2026-05-11 11:27:35 +08:00
1 parent e1088026fa
commit 5e263b6929
5 files changed
+22 -38

No files matched your search

+5 -13
View File
@@ -367,7 +367,7 @@ class OmniDiffusionConfig:
tf_model_config: TransformerConfig = field(default_factory=TransformerConfig)
# Attention
attention_config: "AttentionConfig" = field(default_factory=lambda: AttentionConfig())
diffusion_attention_config: "AttentionConfig" = field(default_factory=lambda: AttentionConfig())
# Running mode
# mode: ExecutionMode = ExecutionMode.INFERENCE
@@ -655,7 +655,7 @@ class OmniDiffusionConfig:
# Match vLLM's config flow: parse entrypoint shorthands before the
# config object is built, and keep a single runtime truth source.
self.attention_config = build_attention_config(self.attention_config)
self.diffusion_attention_config = build_attention_config(self.diffusion_attention_config)
if self.max_cpu_loras is None:
self.max_cpu_loras = 1
@@ -798,20 +798,12 @@ class OmniDiffusionConfig:
else:
kwargs.pop("quantization", None)
# Map "diffusion_attention_config" to "attention_config" so the
# engine_args key (namespaced to avoid collision with vLLM's own
# attention_config) is mapped to the dataclass field.
if "diffusion_attention_config" in kwargs and "attention_config" not in kwargs:
kwargs["attention_config"] = kwargs.pop("diffusion_attention_config")
else:
kwargs.pop("diffusion_attention_config", None)
# Handle "diffusion_attention_backend" shorthand: merge into
# attention_config before field filtering.
# diffusion_attention_config before field filtering.
diffusion_attn_backend = kwargs.pop("diffusion_attention_backend", None)
if diffusion_attn_backend is not None:
existing = kwargs.get("attention_config")
kwargs["attention_config"] = parse_attention_config(
existing = kwargs.get("diffusion_attention_config")
kwargs["diffusion_attention_config"] = parse_attention_config(
existing,
attention_backend=diffusion_attn_backend,
)