biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: reinterpret float32 bits as uint32 before >> 23 for UE8M0

Browse Source
This commit is contained in:
biondizzle
2026-05-11 09:42:03 +00:00
parent f44ff7f6ca
commit 6d7231a50e
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
deep_gemm/mega/__init__.py
View File

@@ -173,10 +173,10 @@ def transform_nvfp4_weights_for_mega_moe(
sf_f32 = sf.to(torch.float32) sf_f32 = sf.to(torch.float32)
# Take max of adjacent pairs # Take max of adjacent pairs
sf_merged = torch.maximum(sf_f32[..., 0::2], sf_f32[..., 1::2]) sf_merged = torch.maximum(sf_f32[..., 0::2], sf_f32[..., 1::2])
# Convert to UE8M0: 2^(floor(log2(v)) - 127 + 127) = extract exponent byte # Convert to UE8M0: extract exponent byte from float32 bit pattern
# UE8M0 encoding: uint8 = float32_exponent_bits >> 23 # UE8M0: uint8 = float32_bits[31:23] (8 exponent bits)
sf_u32 = sf_merged.view(torch.uint32) sf_bits = sf_merged.view(torch.uint32) # reinterpret float32 bits as uint32
sf_ue8m0 = (sf_u32 >> 23).to(torch.uint8) sf_ue8m0 = (sf_bits >> 23).to(torch.uint8)
return sf_ue8m0 return sf_ue8m0
l1_sf_32 = merge_block16_to_block32(l1_sf) l1_sf_32 = merge_block16_to_block32(l1_sf)