chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

chore: fix some typos (#18577)

Browse Source 
Co-authored-by: Liangsheng Yin <lsyincs@gmail.com>
This commit is contained in:
cutetocute
2026-02-11 12:47:41 +08:00
committed by GitHub
parent a2c38f7796
commit 8d2892330c
7 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/attention/double_sparsity_backend.py
View File

@@ -32,7 +32,7 @@ class DoubleSparseAttnBackend(AttentionBackend):
self.heavy_token_num = model_runner.server_args.ds_heavy_token_num
self.sorted_channels = model_runner.sorted_channels
self.sparse_decode_thresold = (
self.sparse_decode_threshold = (
model_runner.server_args.ds_sparse_decode_threshold
)
self.att_out_approx: torch.Tensor = None
@@ -210,7 +210,7 @@ class DoubleSparseAttnBackend(AttentionBackend):
# and set a minimum value for sparse_decode
if (
min_seq_len < self.heavy_token_num
or max_seq_len < self.sparse_decode_thresold
or max_seq_len < self.sparse_decode_threshold
):
self.decode_attention_fwd(
q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),

4
python/sglang/srt/layers/quantization/fp8_kernel.py
View File

@@ -219,7 +219,7 @@ def _per_token_group_quant_8bit_raw(
quantized tensor along with the scaling factor used for quantization.
Args:
x: The input tenosr with ndim >= 2.
x: The input tensor with ndim >= 2.
group_size: The group size used for quantization.
eps: The minimum to avoid dividing zero.
dtype: The dype of output tensor.
@@ -635,7 +635,7 @@ def static_quant_fp8(
quantized tensor along with the scaling factor used for quantization.
Args:
x: The input tenosr with ndim >= 2.
x: The input tensor with ndim >= 2.
x_s: The quantization scale.
repeat_scale: Whether to broadcast per-tensor scale to per-channel scale.
dtype: The dype of output tensor.

2
python/sglang/srt/layers/quantization/int8_kernel.py
View File

@@ -143,7 +143,7 @@ def per_token_group_quant_int8(
quantized tensor along with the scaling factor used for quantization.
Args:
x: The input tenosr with ndim >= 2.
x: The input tensor with ndim >= 2.
group_size: The group size used for quantization.
eps: The minimum to avoid dividing zero.
dtype: The dype of output tensor. Note that only `torch.int8` is supported for now.

4
python/sglang/srt/layers/quantization/quark/utils.py
View File

@@ -128,10 +128,10 @@ def b_dynamic_mxfp4_quant(x):
return x.view(h, b, d // 2), x_scales.view(h, b, d // 32)
def mxfp4_to_f32(x, is_threed):
def mxfp4_to_f32(x, is_3d):
# 2 because we pack fp4 in uint8.
x = x.repeat_interleave(2, dim=-1)
if is_threed:
if is_3d:
x[..., ::2] = x[..., ::2] & 0xF
x[..., 1::2] = x[..., 1::2] >> 4
else:

2
python/sglang/srt/managers/io_struct.py
View File

@@ -1271,7 +1271,7 @@ class UpdateWeightFromDiskReqInput(BaseReq):
torch_empty_cache: bool = False
# Whether to keep the scheduler paused after weight update
keep_pause: bool = False
# Whether to recapture cuda graph after weight udpdate
# Whether to recapture cuda graph after weight update
recapture_cuda_graph: bool = False
# The trainer step id. Used to know which step's weights are used for sampling.
token_step: int = 0