Skip to content

[BugFix] test_mla_fp8.py fails on Cublas 12.9 #11360

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 15 commits into from
Oct 11, 2025
Merged

[BugFix] test_mla_fp8.py fails on Cublas 12.9 #11360

Changes from 4 commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
a2c34a5
modify the fp8quant method for ds2mla head to adapt to bmm_fp8 for
Liu-congo Oct 9, 2025
1729da2
code reformat
Liu-congo Oct 9, 2025
0396a13
remove unused .contiguous() after input_to_float8
Liu-congo Oct 9, 2025
ec5c520
Merge branch 'main' into bmm_fp8_fix
Liu-congo Oct 9, 2025
aa5c275
add special judge for cublas12.9 to maintain performance
Liu-congo Oct 9, 2025
85d989e
Merge branch 'bmm_fp8_fix' of https://github.com/Liu-congo/sglang int…
Liu-congo Oct 9, 2025
e5ffc6f
Merge branch 'main' into bmm_fp8_fix
Liu-congo Oct 9, 2025
d508766
add todo for fixing the triton of per_token_quant_mla_fp8
Liu-congo Oct 9, 2025
f7eec48
Merge branch 'bmm_fp8_fix' of https://github.com/Liu-congo/sglang int…
Liu-congo Oct 9, 2025
2c23687
Merge branch 'main' into bmm_fp8_fix
Liu-congo Oct 10, 2025
5f02def
move version check to srt/utils/common
Liu-congo Oct 10, 2025
b6659d2
Merge branch 'bmm_fp8_fix' of https://github.com/Liu-congo/sglang int…
Liu-congo Oct 10, 2025
e0ecef2
Merge branch 'main' into bmm_fp8_fix
Liu-congo Oct 10, 2025
5f13c02
Merge branch 'main' into bmm_fp8_fix
Fridge003 Oct 10, 2025
d4ee6e0
Merge branch 'main' into bmm_fp8_fix
Fridge003 Oct 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 5 additions & 6 deletions python/sglang/srt/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,6 +94,7 @@
block_quant_dequant,
block_quant_to_tensor_quant,
channel_quant_to_tensor_quant,
input_to_float8,
normalize_e4m3fn_to_e4m3fnuz,
requant_weight_ue8m0_inplace,
)
Expand Down Expand Up @@ -1572,9 +1573,8 @@ def forward_absorb_prepare(
self.w_kc.to(torch.bfloat16) * self.w_scale,
)
elif self.w_kc.dtype == torch.float8_e4m3fn:
q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
q_nope.transpose(0, 1),
zero_allocator.allocate(1),
q_nope_val, q_nope_scale = input_to_float8(
q_nope.transpose(0, 1), torch.float8_e4m3fn
)
q_nope_out = bmm_fp8(
q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
Expand Down Expand Up @@ -1716,9 +1716,8 @@ def forward_absorb_core(
attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)

elif self.w_vc.dtype == torch.float8_e4m3fn:
attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
attn_output.transpose(0, 1),
zero_allocator.allocate(1),
attn_output_val, attn_output_scale = input_to_float8(
attn_output.transpose(0, 1), torch.float8_e4m3fn
)
attn_bmm_output = bmm_fp8(
attn_output_val,
Expand Down
Loading