[3/5][core]refactor communication layer: PR 3 of 5, all other models in non async mode (#3719)

Signed-off-by: natureofnature <wzliu@connect.hku.hk>
This commit is contained in:
NATURE authored and GitHub committed 2026-05-28 14:06:41 +08:00
1 parent 0f3264a9c0
commit 3a347b10b0
38 files changed
+3021 -275

No files matched your search

@@ -59,10 +59,18 @@ class GPUGenerationModelRunner(OmniGPUModelRunner, OmniConnectorModelRunnerMixin
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# Scope full-payload connector init to Qwen3-Omni: other generation
# models (e.g. Bagel DiT) retain their pre-existing runner setup
# so this refactor does not perturb them.
if getattr(self.model_config, "model_arch", None) == "Qwen3OmniMoeForConditionalGeneration":
# Mirrors the init allowlist in gpu_ar_model_runner.py.
_OMNI_CONNECTOR_INIT_ARCHS = {
"Qwen3OmniMoeForConditionalGeneration",
"Qwen2_5OmniForConditionalGeneration",
"CovoAudioForConditionalGeneration",
"MiMoAudioModel",
"Qwen3TTSTalkerForConditionalGeneration",
"Qwen3TTSCode2Wav",
"CosyVoice3Model",
"DyninOmniForConditionalGeneration",
}
if getattr(self.model_config, "model_arch", None) in _OMNI_CONNECTOR_INIT_ARCHS:
self.init_omni_connectors(
vllm_config=self.vllm_config,
model_config=self.model_config,
@@ -299,7 +307,6 @@ class GPUGenerationModelRunner(OmniGPUModelRunner, OmniConnectorModelRunnerMixin
num_tokens_padded,
intermediate_tensors,
)
# [Omni] Pass token counts per request for code2wav output slicing
model_kwargs["seq_token_counts"] = tokens