Rebase 2026/02/23
Update Dockerfile for vllm wheel version and enhance input processing in Omni models. Added decoding for byte messages in async requests, included new metrics in benchmarks, and improved pause state handling in the scheduler. Refined input and output processing methods for better multimodal support and error handling. Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
This commit is contained in:
1 parent
292da4b0d8
commit
bf69cd94bb
9 files changed
+90
-77
No files matched your search
@@ -605,33 +605,35 @@ class GPUGenerationModelRunner(OmniGPUModelRunner):
|
||||
ubatch_slices=ubatch_slices_padded,
|
||||
)
|
||||
|
||||
# If force_attention is True, we always capture attention. Otherwise,
|
||||
# it only happens for cudagraph_runtime_mode=FULL.
|
||||
if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
|
||||
if create_mixed_batch:
|
||||
# In the mixed batch mode (used for FI warmup), we use
|
||||
# shorter sequence lengths to run faster.
|
||||
# TODO(luka) better system for describing dummy batches
|
||||
seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]
|
||||
else:
|
||||
seq_lens = max_query_len # type: ignore[assignment]
|
||||
self.seq_lens.np[:num_reqs] = seq_lens
|
||||
self.seq_lens.np[num_reqs:] = 0
|
||||
self.seq_lens.copy_to_gpu()
|
||||
with self.synchronize_input_prep():
|
||||
# If force_attention is True, we always capture attention.
|
||||
# Otherwise, it only happens for cudagraph_runtime_mode=FULL.
|
||||
if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
|
||||
if create_mixed_batch:
|
||||
# In the mixed batch mode (used for FI warmup), we use
|
||||
# shorter sequence lengths to run faster.
|
||||
# TODO(luka) better system for describing dummy batches
|
||||
seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]
|
||||
else:
|
||||
seq_lens = max_query_len # type: ignore[assignment]
|
||||
self.seq_lens.np[:num_reqs] = seq_lens
|
||||
self.seq_lens.np[num_reqs:] = 0
|
||||
self.seq_lens.copy_to_gpu()
|
||||
|
||||
cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
|
||||
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
|
||||
self.query_start_loc.copy_to_gpu()
|
||||
cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
|
||||
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
|
||||
self.query_start_loc.copy_to_gpu()
|
||||
|
||||
pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
|
||||
attn_metadata, _ = self._build_attention_metadata(
|
||||
num_tokens=num_tokens_unpadded,
|
||||
num_reqs=num_reqs_padded,
|
||||
max_query_len=max_query_len,
|
||||
ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
|
||||
for_cudagraph_capture=is_graph_capturing,
|
||||
slot_mappings=slot_mappings_by_group,
|
||||
)
|
||||
pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
|
||||
attn_metadata, _ = self._build_attention_metadata(
|
||||
num_tokens=num_tokens_unpadded,
|
||||
num_tokens_padded=num_tokens_padded if pad_attn else None,
|
||||
num_reqs=num_reqs_padded,
|
||||
max_query_len=max_query_len,
|
||||
ubatch_slices=(ubatch_slices_padded if pad_attn else ubatch_slices),
|
||||
for_cudagraph_capture=is_graph_capturing,
|
||||
slot_mappings=slot_mappings_by_group,
|
||||
)
|
||||
|
||||
with self.maybe_dummy_run_with_lora(
|
||||
self.lora_config,
|
||||
|
||||
Reference in new issue
Block a user