[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
@@ -26,6 +26,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
||||
)
|
||||
from vllm.models.deepseek_v4.common.ops import fused_indexer_q_rope_quant
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.deep_gemm import mqa_logits_use_mx_scales
|
||||
from vllm.utils.import_utils import has_cutedsl
|
||||
|
||||
HEAD_DIM = 128
|
||||
@@ -347,6 +348,10 @@ def _indexer_k_reference(
|
||||
exponent = k.float().abs().amax(dim=-1, keepdim=True).clamp(min=1e-4) / 448
|
||||
scale = exponent.log2().ceil().exp2()
|
||||
values = (k.float() / scale).clamp(-448, 448).to(torch.float8_e4m3fn)
|
||||
if mqa_logits_use_mx_scales():
|
||||
# The power-of-two scale's exponent as 4 equal UE8M0 bytes.
|
||||
ue8m0 = (scale.view(torch.int32) >> 23).to(torch.uint8)
|
||||
return values.view(torch.uint8), ue8m0.repeat(1, 4)
|
||||
return values.view(torch.uint8), scale.view(torch.uint8)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user