Signed-off-by: Jialin Ouyang <Jialin.Ouyang@gmail.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: Jialin Ouyang <Jialin.Ouyang@gmail.com>
This commit is contained in:
Cyrus Leung
2025-11-14 22:47:41 -08:00
committed by GitHub
co-authored by Jialin Ouyang
parent 6965ef436f
commit 98b4d389ed
15 changed files with 122 additions and 91 deletions
+23 -13
View File
@@ -216,9 +216,11 @@ class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
del self._logprobs_tensors
del self._sampled_token_ids
valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
valid_sampled_token_ids: list[np.ndarray] = [
row for row in self.sampled_token_ids_cpu.numpy()
]
for i in self._invalid_req_indices:
valid_sampled_token_ids[i].clear()
valid_sampled_token_ids[i] = np.array([])
output = self._model_runner_output
output.sampled_token_ids = valid_sampled_token_ids
@@ -2339,7 +2341,7 @@ class GPUModelRunner(
) -> tuple[
dict[str, int],
LogprobsLists | None,
list[list[int]],
list[np.ndarray],
dict[str, LogprobsTensors | None],
list[str],
dict[str, int],
@@ -2365,6 +2367,7 @@ class GPUModelRunner(
num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
sampled_token_ids = sampler_output.sampled_token_ids
invalid_req_indices = []
valid_sampled_token_ids: list[np.ndarray]
if not self.use_async_scheduling:
# Get the valid generated tokens.
max_gen_len = sampled_token_ids.shape[-1]
@@ -2379,7 +2382,7 @@ class GPUModelRunner(
)
# Mask out the sampled tokens that should not be sampled.
for i in discard_sampled_tokens_req_indices:
valid_sampled_token_ids[int(i)].clear()
valid_sampled_token_ids[int(i)] = np.array([])
else:
valid_sampled_token_ids = []
invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
@@ -2407,19 +2410,24 @@ class GPUModelRunner(
[0] if spec_decode_metadata and logprobs_tensors else None
)
for req_idx in range(num_sampled_tokens):
sampled_ids: np.ndarray | None
if self.use_async_scheduling:
sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
sampled_ids = (
np.array([-1]) if req_idx not in invalid_req_indices_set else None
)
else:
sampled_ids = valid_sampled_token_ids[req_idx]
num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
num_sampled_ids: int = (
sampled_ids.shape[0] if sampled_ids is not None else 0
)
if cu_num_accepted_tokens is not None:
cu_num_accepted_tokens.append(
cu_num_accepted_tokens[-1] + num_sampled_ids
)
if not sampled_ids:
if sampled_ids is None or num_sampled_ids == 0:
continue
start_idx = self.input_batch.num_tokens_no_spec[req_idx]
@@ -2761,7 +2769,9 @@ class GPUModelRunner(
with record_function_or_nullcontext("gpu_model_runner: sample"):
sampler_output = self._sample(logits, spec_decode_metadata)
def propose_draft_token_ids(sampled_token_ids):
def propose_draft_token_ids(
sampled_token_ids: torch.Tensor | list[np.ndarray],
) -> None:
assert spec_decode_common_attn_metadata is not None
with record_function_or_nullcontext("gpu_model_runner: draft"):
self._draft_token_ids = self.propose_draft_token_ids(
@@ -2883,14 +2893,14 @@ class GPUModelRunner(
def propose_draft_token_ids(
self,
scheduler_output: "SchedulerOutput",
sampled_token_ids: torch.Tensor | list[list[int]],
sampled_token_ids: torch.Tensor | list[np.ndarray],
sampling_metadata: SamplingMetadata,
hidden_states: torch.Tensor,
sample_hidden_states: torch.Tensor,
aux_hidden_states: list[torch.Tensor] | None,
spec_decode_metadata: SpecDecodeMetadata | None,
common_attn_metadata: CommonAttentionMetadata,
) -> list[list[int]] | torch.Tensor:
) -> torch.Tensor | list[list[int]]:
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
if self.speculative_config.method == "ngram":
assert isinstance(sampled_token_ids, list)
@@ -2922,7 +2932,7 @@ class GPUModelRunner(
for num_draft, tokens in zip(
spec_decode_metadata.num_draft_tokens, sampled_token_ids
):
indices.append(offset + len(tokens) - 1)
indices.append(offset + tokens.shape[0] - 1)
offset += num_draft + 1
indices = torch.tensor(indices, device=self.device)
hidden_states = sample_hidden_states[indices]
@@ -4862,7 +4872,7 @@ class GPUModelRunner(
return kv_cache_spec
def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
def _to_list(self, sampled_token_ids: torch.Tensor) -> list[np.ndarray]:
# This is a short term mitigation for issue mentioned in
# https://github.com/vllm-project/vllm/issues/22754.
# `tolist` would trigger a cuda wise stream sync, which
@@ -4875,4 +4885,4 @@ class GPUModelRunner(
pinned.copy_(sampled_token_ids, non_blocking=True)
self.transfer_event.record()
self.transfer_event.synchronize()
return pinned.tolist()
return [row for row in pinned.numpy()]