rebase: align vllm-omni with vLLM e9163b536e72

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
This commit is contained in:
tzhouam committed 2026-03-16 02:32:39 +00:00
1 parent f137a1e3ad
commit 3a508d7b7a
14 files changed
+97 -46

No files matched your search

@@ -87,7 +87,7 @@ class GPUGenerationModelRunner(OmniGPUModelRunner):
if self.execute_model_state is not None:
raise RuntimeError("State error: sample_tokens() must be called after execute_model() returns None.")
if self.vllm_config.model_config.enable_return_routed_experts:
if self.routed_experts_initialized:
capturer = RoutedExpertsCapturer.get_instance()
if capturer is not None:
capturer.clear_buffer() # noqa
@@ -265,9 +265,9 @@ class GPUGenerationModelRunner(OmniGPUModelRunner):
# Run the model.
# Use persistent buffers for CUDA graphs.
# When spec decode is enabled, defer clearing KV connector metadata
# until after draft model runs in sample_tokens.
clear_kv_metadata = self.speculative_config is None
# When spec decode is enabled, defer connector finalization
# (wait_for_save + clear metadata) until after draft model runs.
defer_kv_connector_finalize = self.speculative_config is not None
with (
set_forward_context(
attn_metadata,
@@ -282,7 +282,7 @@ class GPUGenerationModelRunner(OmniGPUModelRunner):
record_function_or_nullcontext("Forward"),
self.maybe_get_kv_connector_output(
scheduler_output,
clear_metadata=clear_kv_metadata,
defer_finalize=defer_kv_connector_finalize,
) as kv_connector_output,
):
outputs = self._run_generation_model(