[Kernel][Build] Bump DeepGEMM to the 26/09/30 release and adapt the sparse indexer (#61149)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
86b1c2ef15
commit
a6009ea85b
@@ -29,6 +29,7 @@ import pytest
|
||||
import torch
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.deep_gemm import mqa_logits_use_mx_scales
|
||||
|
||||
if current_platform.is_rocm():
|
||||
from vllm.models.glm5next.amd.ops.kpool_compress import (
|
||||
@@ -206,7 +207,12 @@ def _torch_reference(
|
||||
k_off = page_base + loc_tok * HEAD_DIM + torch.arange(HEAD_DIM)
|
||||
s_off = page_base + k_region + loc_tok * 4
|
||||
kv_flat[k_off] = quantized.view(torch.uint8)
|
||||
kv_flat[s_off : s_off + 4] = scale.detach().reshape(1).view(torch.uint8)
|
||||
scale_bytes = scale.detach().reshape(1).view(torch.uint8)
|
||||
if mqa_logits_use_mx_scales():
|
||||
# The power-of-two scale's exponent as 4 equal UE8M0 bytes.
|
||||
ue8m0 = scale.reshape(1).view(torch.int32) >> 23
|
||||
scale_bytes = ue8m0.to(torch.uint8).repeat(4)
|
||||
kv_flat[s_off : s_off + 4] = scale_bytes
|
||||
|
||||
# stash -- gated on the TOKEN-granular tail slot, not on pos_valid.
|
||||
# pos_valid keys off the pool-granular cache_loc, which is -1 for
|
||||
|
||||
Reference in New Issue
Block a user