xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-04-12 10:47:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

tweaks

Browse Source 
Signed-off-by: Tyler Michael Smith <tysmith@redhat.com>
This commit is contained in:
Tyler Michael Smith 2025-06-20 15:36:59 +00:00
parent 7a821f0e7f
commit 39d5d33f8f
1 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py
View File

@ -17,14 +17,12 @@ has_deep_gemm = importlib.util.find_spec("deep_gemm") is not None
@triton.jit
def _silu_mul_fp8_quant_deep_gemm(
# Pointers ------------------------------------------------------------
input_ptr, # *FP32 activations (E, T, 2*H)
y_q_ptr, # *FP8 quantised activations (E, T, H)
y_s_ptr, # *FP32 scales (E, T, G)
counts_ptr, # *INT32 number of tokens per expert (E)
input_ptr, # 16-bit activations (E, T, 2*H)
y_q_ptr, # fp88 quantized activations (E, T, H)
y_s_ptr, # 16-bit scales (E, T, G)
counts_ptr, # int32 num tokens per expert (E)
# Sizes ---------------------------------------------------------------
E: tl.constexpr, # num_experts
T: tl.constexpr, # max_num_tokens
H: tl.constexpr, # hidden dimension (per output)
GROUP_SIZE: tl.constexpr, # elements per group (usually 128)
@ -159,8 +157,6 @@ def silu_mul_fp8_quant_deep_gemm(
y_q,
y_s,
tokens_per_expert,
E,
T,
H,
group_size,
stride_i_e,