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

[AMD] Fix DeepSeek block-FP8 loading on gfx94x (#38446)

Browse Source 
Signed-off-by: Bingxu Chen <bingxche@amd.com>
This commit is contained in:
Bingxu Chen
2026-09-10 23:00:29 -07:00
committed by GitHub
parent b9899b04c1
commit 2465ee3948
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_common/deepseek_weight_loader.py
+5 -2
View File
@@ -628,7 +628,10 @@ class DeepseekV2WeightLoaderMixin:
if hasattr(self_attn.kv_b_proj, "weight_scale") if hasattr(self_attn.kv_b_proj, "weight_scale")
else self_attn.kv_b_proj.weight_scale_inv else self_attn.kv_b_proj.weight_scale_inv
) )
if _is_fp8_fnuz: is_ue8m0_uint8 = (
weight_scale.format_ue8m0 and weight_scale.dtype == torch.uint8
)
if _is_fp8_fnuz and not is_ue8m0_uint8:
weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz( weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
weight=w, weight=w,
weight_scale=weight_scale, weight_scale=weight_scale,
@@ -638,7 +641,7 @@ class DeepseekV2WeightLoaderMixin:
weight = w weight = w
# In multiple weight loading scenarios (e.g. RL), we need to inverse the scale of the weights after the requantization happened at the first loading. # In multiple weight loading scenarios (e.g. RL), we need to inverse the scale of the weights after the requantization happened at the first loading.
if weight_scale.format_ue8m0 and weight_scale.dtype == torch.uint8: if is_ue8m0_uint8:
weight_scale = (weight_scale.to(torch.int32) << 23).view( weight_scale = (weight_scale.to(torch.int32) << 23).view(
torch.float32 torch.float32
) )