[Doc] Sync four function docstrings with their signatures (#45437)
Signed-off-by: Daoyuan Li <94409450+DaoyuanLi2816@users.noreply.github.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Daoyuan Li committed
Jul 14, 2026 at 10:10 UTC
ca3618bc698ed029c971a359f9b2b5bc1d041bb8
3 files changed
+8
-4
vllm/_aiter_ops.py
+2
-2
@@ -2637,7 +2637,8 @@ class rocm_aiter_ops:
2637
2638
Args:
2639
tensor: The input weight tensor to be shuffled.
2640
- layout: The block layout to use, defaults to (16, 4).
2640
+ nLane: Number of lanes in the shuffle layout.
2641
+ gate_up: Whether the weight is for w13 (True) or w2 (False).
2642
2643
Returns:
2644
torch.Tensor: The shuffled tensor.
@@ -2875,7 +2876,6 @@ class rocm_aiter_ops:
2876
hc_sinkhorn_eps: sinkhorn epsilon
2877
hc_post_mult_value: post-mix multiplier value
2878
sinkhorn_repeat: number of sinkhorn iterations
2878
- n_splits: split-k factor;
2879
2880
Returns:
2881
post_mix: shape (..., hc_mult), dtype torch.float32
vllm/_custom_ops.py
+5
-1
@@ -1539,7 +1539,11 @@ def scaled_fp4_quant(
1539
Args:
1540
input: The input tensor to be quantized to FP4
1541
input_global_scale: A scalar scaling factor for the entire tensor.
1542
- use_8x4_sf_layout: Whether to use the 8x4 or 128x4 layout for the scaling
1542
+ is_sf_swizzled_layout: Whether to store the scaling factors in the
1543
+ swizzled layout (default `True`).
1544
+ backend: Quantization kernel backend to dispatch to. For `"trtllm"`
1545
+ backends the 8x4 scale-factor layout is selected for small
1546
+ batches (m <= 32) instead of the 128x4 layout.
1547
padded_n: Optional padded K dimension. When provided, the quantized
1548
output and scale tensors are allocated for ``padded_n``
1549
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+1
-1
@@ -909,7 +909,7 @@ def w8a8_triton_block_scaled_mm(
909
Bs: The per-block quantization scale for `B`.
910
block_size: The block size for per-block quantization. It should
911
be 2-dim, e.g., [128, 128].
912
- output_dytpe: The dtype of the returned tensor.
912
+ output_dtype: The dtype of the returned tensor.
913
Returns:
914
torch.Tensor: The result of matmul.
915
"""