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
@@ -91,7 +91,8 @@ def validate_image_sentinel_ids(tokenizer) -> None:
|
||||
|
||||
def grid_tokens(best_height, best_width, patch_size, downsample_ratio):
|
||||
"""Number of LLM tokens the aligner grid occupies (N-layout, including
|
||||
row/align padding)."""
|
||||
row/align padding).
|
||||
"""
|
||||
n_llm_h = math.ceil((best_height // patch_size) / downsample_ratio)
|
||||
n_llm_w = math.ceil((best_width // patch_size) / downsample_ratio)
|
||||
num_tokens = n_llm_h * (n_llm_w + 1) + 2
|
||||
@@ -196,7 +197,8 @@ def load_image(
|
||||
|
||||
def build_image_block(n_llm_h: int, n_llm_w: int, start_pos: int):
|
||||
"""Builds the N-layout token types (final order) and the aligner-row order
|
||||
for IMAGE slots."""
|
||||
for IMAGE slots.
|
||||
"""
|
||||
compress_pad = COMPRESS_PAD_TO - 1 - start_pos % COMPRESS_PAD_TO
|
||||
pad_h = n_llm_h % 2
|
||||
rows = n_llm_h + pad_h
|
||||
@@ -239,7 +241,8 @@ def build_image_block_pad_free(n_llm_h: int, n_llm_w: int):
|
||||
|
||||
class DeepseekV4VLImageProcessor:
|
||||
"""Per-image transform (the PIL-input equivalent of the reference
|
||||
``load_image``)."""
|
||||
``load_image``).
|
||||
"""
|
||||
|
||||
def __init__(self, config: DeepseekV4Config) -> None:
|
||||
super().__init__()
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Triton kernels for DeepseekV4 paged K-cache management and sparse-attention index
|
||||
"""Triton kernels for DeepseekV4 paged K-cache management and sparse-attention index
|
||||
preparation.
|
||||
|
||||
- quantize_and_insert_k_cache: quantize bf16 K to UE8M0 FP8 and insert into
|
||||
@@ -61,8 +60,7 @@ def quantize_and_insert_k_kernel(
|
||||
n_quant_blocks: tl.constexpr, # 8 (7 real + 1 padding)
|
||||
use_fnuz: tl.constexpr = False,
|
||||
):
|
||||
"""
|
||||
Quantize K tensor and insert into paged K cache.
|
||||
"""Quantize K tensor and insert into paged K cache.
|
||||
|
||||
K Cache block layout (block_size=64 tokens):
|
||||
- [0, 64*576): Token data, each token has 448 fp8 + 128 bf16
|
||||
@@ -174,8 +172,7 @@ def quantize_and_insert_k_cache(
|
||||
is_ue8m0: bool = True,
|
||||
use_fnuz: bool = False,
|
||||
):
|
||||
"""
|
||||
Quantize K tensor and insert into paged K cache.
|
||||
"""Quantize K tensor and insert into paged K cache.
|
||||
|
||||
K Cache block layout (block_size=64 tokens):
|
||||
- First 64 * 576 = 36864 bytes: Token data
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Fused compressor + FP8/MXFP4 UE8M0 quantization + KV cache insert kernels.
|
||||
"""Fused compressor + FP8/MXFP4 UE8M0 quantization + KV cache insert kernels.
|
||||
|
||||
Three specialized kernels:
|
||||
- _fused_kv_compress_norm_rope_insert_sparse_attn:
|
||||
@@ -390,7 +389,7 @@ def _compress_gather_split_sparse_attn(
|
||||
NUM_SPLITS: tl.constexpr,
|
||||
HEAD_TILE: tl.constexpr, # HEAD_SIZE // NUM_SPLITS
|
||||
):
|
||||
"""Stage 1: per-(token, head-split) compress gather, write to fp32 scratch
|
||||
"""Stage 1: per-(token, head-split) compress gather, write to fp32 scratch.
|
||||
|
||||
No-overlap gather (cr>=128) on rows [0, COMPRESS_RATIO)
|
||||
"""
|
||||
@@ -639,7 +638,7 @@ def compress_norm_rope_store_two_stage_triton(
|
||||
num_decode_tokens: int,
|
||||
compress_scratch: torch.Tensor,
|
||||
) -> None:
|
||||
"""Two-stage split compressor dispatch for head=512 cr>=128 (no-overlap)
|
||||
"""Two-stage split compressor dispatch for head=512 cr>=128 (no-overlap).
|
||||
|
||||
Run the occupancy-fanned two-stage split for prefill [num_decodee_tokens:]
|
||||
to fill the CUs, and use the original single-pass launcher
|
||||
|
||||
@@ -312,7 +312,8 @@ class FusedIndexerQRopeQuantTritonKernel(
|
||||
def _indexer_weights_out_dtypes(vllm_config: Any) -> tuple[torch.dtype, ...]:
|
||||
"""Weights dtypes the model's indexer layers ask for: fp32 for the dense
|
||||
scoring kernels, plus bf16 when the DeepSeek V4.1 sparse-logits indexer
|
||||
(`SparseMQAIndexer`) is enabled."""
|
||||
(`SparseMQAIndexer`) is enabled.
|
||||
"""
|
||||
if vllm_config.attention_config.indexer_sparse_logits:
|
||||
return (torch.float32, torch.bfloat16)
|
||||
return (torch.float32,)
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Fused inverse RoPE + block-scaled FP8 quantization kernel for DeepseekV4 attention.
|
||||
"""Fused inverse RoPE + block-scaled FP8 quantization kernel for DeepseekV4 attention.
|
||||
|
||||
Output scale format is pre-transformed (MN-major TMA-aligned; FP32 on SM90,
|
||||
INT32-packed UE8M0 on SM100) so fp8_einsum skips transform_sf_into_required_layout.
|
||||
@@ -380,6 +379,7 @@ def fused_inv_rope_fp8_quant(
|
||||
Returns:
|
||||
Rotated output in [T, G, D] and its FP8 scales. The scale tensor is
|
||||
empty when quantization is disabled.
|
||||
|
||||
"""
|
||||
from vllm.utils.deep_gemm import get_tma_aligned_size
|
||||
|
||||
|
||||
@@ -263,6 +263,7 @@ def run_dp_sharded_vision_tower(
|
||||
|
||||
Returns:
|
||||
One ``(n_aligner_rows, hidden_size)`` embedding tensor per image.
|
||||
|
||||
"""
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
|
||||
Reference in New Issue
Block a user