[Core] Check for GPU<->CPU syncs during CI (#43107)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nick Hill
2026-08-14 17:42:05 -07:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 7b544ecb52
commit 615d4cfade
44 changed files with 600 additions and 357 deletions
+26 -20
View File
@@ -120,6 +120,7 @@ from vllm.sequence import IntermediateTensors
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
from vllm.tracing import instrument
from vllm.utils import length_from_prompt_token_ids_or_embeds
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
from vllm.utils.math_utils import cdiv, round_up
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
from vllm.utils.nvtx_pytorch_hooks import PytHooks
@@ -3848,26 +3849,28 @@ class GPUModelRunner(
self.routed_experts_slot_mapping_device[:total],
non_blocking=True,
)
with gpu_sync_allowed():
# Get the valid generated tokens.
max_gen_len = sampled_token_ids.shape[-1]
if max_gen_len == 1:
# No spec decode tokens.
valid_sampled_token_ids = self._to_list(sampled_token_ids)
# 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()
# Get the valid generated tokens.
max_gen_len = sampled_token_ids.shape[-1]
if max_gen_len == 1:
# No spec decode tokens.
valid_sampled_token_ids = self._to_list(sampled_token_ids)
# 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()
if logprobs_tensors is not None:
logprobs_lists = logprobs_tensors.tolists()
else:
# Includes spec decode tokens.
valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
if logprobs_tensors is not None:
logprobs_lists = logprobs_tensors.tolists()
else:
# Includes spec decode tokens.
valid_sampled_token_ids, logprobs_lists = (
RejectionSampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
)
else:
valid_sampled_token_ids = []
invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
@@ -5843,7 +5846,10 @@ class GPUModelRunner(
on device instead; see`AsyncGPUModelRunnerOutput`.
"""
try:
counts = [] if logits is None else count_nans_per_row(logits).tolist()
# Reporting per-request NaN counts requires them on the host; this
# path is opt-in diagnostics, so the D2H is intended.
with gpu_sync_allowed():
counts = [] if logits is None else count_nans_per_row(logits).tolist()
num_nans_in_logits = nans_to_dict(counts, self.input_batch.req_id_to_index)
if envs.VLLM_RAISE_ON_LOGIT_NANS:
raise_if_nan_logits(num_nans_in_logits)