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
@@ -130,7 +130,8 @@ def _resolve_kda_prefill_backend(
|
||||
"""Pick the chunked-prefill kernel: FlashKDA (fused CUDA, ~2-4x faster on
|
||||
SM90/SM10x/SM12x for bf16, head_dim 128 and a bounded gate) or the Triton
|
||||
``chunk_kda_with_fused_gate`` path. ``backend`` comes from
|
||||
``additional_config.kda_prefill_backend`` (auto / triton / flashkda)."""
|
||||
``additional_config.kda_prefill_backend`` (auto / triton / flashkda).
|
||||
"""
|
||||
if backend not in ("auto", "triton", "flashkda"):
|
||||
raise ValueError(f"Unsupported KDA prefill backend: {backend}")
|
||||
capability = current_platform.get_device_capability()
|
||||
@@ -368,7 +369,8 @@ class Glm5NextLinearAttention(GatedDeltaNetAttention):
|
||||
bounded gate ``lower_bound * sigmoid(exp(A_log) * (g + dt_bias))``,
|
||||
matching ``chunk_kda_with_fused_gate(..., safe_gate=True)``. Writes the
|
||||
attention output into ``out`` (a workspace buffer when ``None``) and
|
||||
returns ``(out, final_state)``."""
|
||||
returns ``(out, final_state)``.
|
||||
"""
|
||||
assert self._flashkda_buffer_specs is not None
|
||||
final_state, workspace, workspace_out = (
|
||||
current_workspace_manager().get_simultaneous(*self._flashkda_buffer_specs)
|
||||
|
||||
@@ -720,7 +720,8 @@ class Glm5NextMultiModalProcessor(Glm4vMultiModalProcessor):
|
||||
(image token repeat, video frame/timestamp structure) is owned by vLLM's
|
||||
prompt-update machinery — the inherited ``_get_prompt_updates`` builds
|
||||
the replacement content and the placeholder scan validates against
|
||||
exactly that."""
|
||||
exactly that.
|
||||
"""
|
||||
|
||||
def _hf_processor_applies_updates(
|
||||
self,
|
||||
|
||||
@@ -138,7 +138,8 @@ def _gather_workspace_shapes(
|
||||
"""Return ((values_shape, values_dtype), (scales_shape, scales_dtype)) for
|
||||
the K-gather workspace. FP8 path: (T, head_dim) fp8 + (T, 4) uint8 fp32
|
||||
scales. MXFP4 path: (T, head_dim // 2) uint8 packed mxfp4 +
|
||||
(T, head_dim // MXFP4_BLOCK_SIZE) uint8 ue8m0 scales."""
|
||||
(T, head_dim // MXFP4_BLOCK_SIZE) uint8 ue8m0 scales.
|
||||
"""
|
||||
if use_fp4_cache:
|
||||
return (
|
||||
((total_seq_lens, head_dim // 2), torch.uint8),
|
||||
@@ -156,7 +157,8 @@ def kv_cache_as_quant_view(
|
||||
use_fp4_cache: bool,
|
||||
) -> torch.Tensor:
|
||||
"""4D ``[num_blocks, block_size, 1, head_width]`` view expected by
|
||||
DeepGEMM, from the 3D indexer kv-cache allocation."""
|
||||
DeepGEMM, from the 3D indexer kv-cache allocation.
|
||||
"""
|
||||
if use_fp4_cache:
|
||||
assert kv_cache.ndim == 3 and kv_cache.dtype == torch.uint8
|
||||
num_blocks, block_size, _ = kv_cache.shape
|
||||
|
||||
Reference in New Issue
Block a user