Add Qwen3.6 MoE inference support

2026-07-13 20:24:41 +08:00
parent 588bfd9df3
commit a2de146fb6
27 changed files with 3153 additions and 149 deletions
--- a/csrc/attention/paged_attention.cu
+++ b/csrc/attention/paged_attention.cu
@@ -21,11 +21,11 @@
 //                 active batch) so we just index by blockIdx.x_seq.
 //   context_lens  [batch] int32  — number of valid tokens per sequence.
 //
-// One CUDA block: 256 threads, head_dim <= 128.
+// One CUDA block: 256 threads, head_dim <= 256.

 #define PAGED_BLOCK_SIZE 16
 #define PAGED_THREADS 256
-#define PAGED_HEAD_DIM_MAX 128
+#define PAGED_HEAD_DIM_MAX 256

 __global__ void paged_decode_attention_bf16_kernel(
    const __nv_bfloat16* __restrict__ Q,