Revert "[V1] Exception Handling when Loading KV Cache from Remote Store" (#21778)

Signed-off-by: KuntaiDu <kuntai@uchicago.edu>
This commit is contained in:
Kuntai Du
2025-07-28 20:15:18 +00:00
committed by GitHub
parent 9ba1c88a93
commit b18b417fbf
10 changed files with 5 additions and 229 deletions
+2 -8
View File
@@ -1375,7 +1375,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
num_scheduled_tokens_np: np.ndarray,
finished_sending: Optional[set[str]],
finished_recving: Optional[set[str]],
finished_loading_dict: Optional[dict[str, int]],
) -> ModelRunnerOutput:
assert self.input_batch.num_reqs ==\
len(self.input_batch.pooling_params), \
@@ -1412,7 +1411,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
pooler_output=pooler_output,
finished_sending=finished_sending,
finished_recving=finished_recving,
finished_loading_dict=finished_loading_dict,
)
@torch.inference_mode()
@@ -1532,7 +1530,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
self.maybe_wait_for_kv_save()
finished_sending, finished_recving = (
self.get_finished_kv_transfers(scheduler_output))
finished_loading_dict = self.get_finished_loading(scheduler_output)
if self.use_aux_hidden_state_outputs:
hidden_states, aux_hidden_states = model_output
@@ -1550,11 +1547,9 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
if not get_pp_group().is_last_rank:
# For mid-pipeline stages, return the hidden states.
if not broadcast_pp_output:
if (finished_sending or finished_recving
or finished_loading_dict):
if finished_sending or finished_recving:
hidden_states.finished_sending = finished_sending
hidden_states.finished_recving = finished_recving
hidden_states.finished_loading_dict = finished_loading_dict
return hidden_states
assert isinstance(hidden_states, IntermediateTensors)
get_pp_group().send_tensor_dict(hidden_states.tensors,
@@ -1564,7 +1559,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
if self.input_batch.pooling_params:
return self._pool(hidden_states, num_scheduled_tokens,
num_scheduled_tokens_np, finished_sending,
finished_recving, finished_loading_dict)
finished_recving)
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states, None)
@@ -1716,7 +1711,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
pooler_output=[],
finished_sending=finished_sending,
finished_recving=finished_recving,
finished_loading_dict=finished_loading_dict,
num_nans_in_logits=num_nans_in_logits,
)