【Fix bug] w4afp8 的nblock固定为256，并且fa3的append attn 增加mask参数 (#3771) (#3835)

* fix w4afp8

* 增加集中式配置

* codestyle

* fix fa3 append attn

custom_ops/gpu_ops/w4afp8_gemm/w4afp8_gemm.cu

@@ -75,12 +75,8 @@ void DisPatchW4AFp8Gemm(
         const int64_t K,
         cudaStream_t stream) {
 
-    int kBlockN = (max_tokens + 15) / 16 * 16;
+    int kBlockN = 256;
     int TailN = 0;
-    if (kBlockN > 256) {
-        TailN = kBlockN % 256;
-        kBlockN = 256;
-    }
     if constexpr (std::is_same_v<OutputType, cutlass::bfloat16_t>) {
         GEMM_SWITCH_BF16(
             M, K, batch_size, token_padding_size, kBlockN, TailN,

custom_ops/utils/auto_gen_w4afp8_gemm_kernel.py

@@ -88,6 +88,8 @@ gemm_case = [
     [8192, 3584, 8, 2048],  # eb45T ffn1
     [7168, 8192, 8, 0],  # eb45T ffn2
     [7168, 8192, 8, 2048],  # eb45T ffn2
+    [1792, 8192, 64, 0],  # eb45t ffn1
+    [8192, 896, 64, 0],  # eb45t ffn2
 ]
 
 dtype = ["BF16"]

fastdeploy/model_executor/layers/attention/flash_attn_backend.py

@@ -359,6 +359,7 @@ class FlashAttentionBackend(AttentionBackend):
             getattr(layer, "cache_v_zp", None),
             layer.linear_shift,
             layer.linear_smooth,
+            forward_meta.attn_mask_offsets,
             metadata.kv_signal_data_list[layer.layer_id],
             getattr(layer, "q_norm_weight", None),
             getattr(layer, "k_norm_weight", None),