Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
GPU-accelerated N-gram proposer using fully async PyTorch tensor operations.
|
||||
"""GPU-accelerated N-gram proposer using fully async PyTorch tensor operations.
|
||||
|
||||
This version uses a fully vectorized approach with unfold and argmax for
|
||||
finding the first match across all sequences in parallel.
|
||||
@@ -52,8 +51,7 @@ class NgramGPUKernel(nn.Module):
|
||||
max_ngram_len: int,
|
||||
num_draft_tokens: int,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Find suffix n-gram matches and extract following tokens.
|
||||
"""Find suffix n-gram matches and extract following tokens.
|
||||
Searches for the earliest prior occurrence of the trailing n-gram,
|
||||
tries multiple lengths, and picks the longest valid match.
|
||||
|
||||
@@ -66,6 +64,7 @@ class NgramGPUKernel(nn.Module):
|
||||
|
||||
Returns:
|
||||
Draft token predictions; -1 means invalid/no match.
|
||||
|
||||
"""
|
||||
batch_size = token_ids.shape[0]
|
||||
max_seq_len = token_ids.shape[1]
|
||||
@@ -165,8 +164,7 @@ class NgramGPUKernel(nn.Module):
|
||||
token_ids_gpu: torch.Tensor,
|
||||
combined_mask: torch.Tensor,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Forward pass for N-gram proposal using GPU tensor operations.
|
||||
"""Forward pass for N-gram proposal using GPU tensor operations.
|
||||
|
||||
Args:
|
||||
num_tokens_no_spec: Number of tokens for each sequence [batch_size]
|
||||
@@ -177,8 +175,8 @@ class NgramGPUKernel(nn.Module):
|
||||
draft_tokens: [batch_size, k] on GPU
|
||||
num_valid_draft_tokens: [batch_size] int32 on GPU, count of
|
||||
leading valid (non -1) tokens per request.
|
||||
"""
|
||||
|
||||
"""
|
||||
device = token_ids_gpu.device
|
||||
|
||||
# Infer batch size to preserve dynamic shape.
|
||||
@@ -282,8 +280,7 @@ class NgramProposerGPU:
|
||||
pattern_len: int,
|
||||
device: str = "cuda",
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Generate random test data with n-gram repetitions.
|
||||
"""Generate random test data with n-gram repetitions.
|
||||
|
||||
Args:
|
||||
batch_size: Number of sequences in the batch
|
||||
@@ -296,6 +293,7 @@ class NgramProposerGPU:
|
||||
num_tokens: [batch_size] tensor
|
||||
sampled_flags: [batch_size] bool tensor
|
||||
valid_mask: [batch_size] bool tensor
|
||||
|
||||
"""
|
||||
token_ids = torch.zeros(
|
||||
batch_size,
|
||||
@@ -321,8 +319,7 @@ class NgramProposerGPU:
|
||||
valid_sampled_token_ids_gpu: torch.Tensor, # [batch_size, num_spec_tokens + 1]
|
||||
valid_sampled_tokens_count: torch.Tensor, # [batch_size]
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Propose draft tokens using GPU-accelerated n-gram matching.
|
||||
"""Propose draft tokens using GPU-accelerated n-gram matching.
|
||||
|
||||
Scatter sampled tokens into `token_ids_gpu`, compute temporary
|
||||
updated lengths, then run the kernel.
|
||||
@@ -338,6 +335,7 @@ class NgramProposerGPU:
|
||||
draft_tokens: Proposed draft token IDs [batch_size, k]
|
||||
num_valid_draft_tokens: Count of leading valid draft tokens
|
||||
per request [batch_size]
|
||||
|
||||
"""
|
||||
assert num_speculative_tokens == self.k
|
||||
assert token_ids_gpu.device == self.device
|
||||
@@ -398,8 +396,7 @@ class NgramProposerGPU:
|
||||
num_tokens_no_spec: torch.Tensor,
|
||||
discard_request_mask: torch.Tensor,
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Prepare speculative decoding inputs on device:
|
||||
"""Prepare speculative decoding inputs on device:
|
||||
compute next token ids and valid counts, honoring discarded requests
|
||||
and rejected tokens, without CPU-GPU sync.
|
||||
"""
|
||||
@@ -485,6 +482,7 @@ def update_scheduler_for_invalid_drafts(
|
||||
num_valid_draft_tokens_cpu: CPU buffer of valid draft counts.
|
||||
scheduler_output: Scheduler metadata to update in-place.
|
||||
req_id_to_index: Request-id to batch-index mapping.
|
||||
|
||||
"""
|
||||
req_data = scheduler_output.scheduled_cached_reqs
|
||||
num_valid_draft_tokens_event.synchronize()
|
||||
@@ -652,9 +650,7 @@ def copy_num_valid_draft_tokens(
|
||||
num_valid_draft_tokens: torch.Tensor | None,
|
||||
batch_size: int,
|
||||
) -> None:
|
||||
"""
|
||||
Async D2H copy of per-request valid draft counts.
|
||||
"""
|
||||
"""Async D2H copy of per-request valid draft counts."""
|
||||
if num_valid_draft_tokens is None:
|
||||
return
|
||||
|
||||
|
||||
Reference in New Issue
Block a user