[Spec] Remove dead kernel params; fix stale comment in trtllm_mla (#25010)

python/sglang/srt/layers/attention/fla/kda.py

@@ -52,7 +52,6 @@ def fused_recurrent_kda_fwd(
     inplace_final_state: bool = True,
     cu_seqlens: torch.LongTensor | None = None,
     # ssm_state_indices: torch.Tensor | None = None,
-    num_accepted_tokens: torch.Tensor | None = None,
     use_qk_l2norm_in_kernel: bool = False,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     B, T, H, K, V = *k.shape, v.shape[-1]
@@ -92,7 +91,6 @@ def fused_recurrent_kda_fwd(
         ht=final_state,
         cu_seqlens=cu_seqlens,
         # ssm_state_indices=ssm_state_indices,
-        # num_accepted_tokens=num_accepted_tokens,
         scale=scale,
         # N=N,
         T=T,
@@ -155,7 +153,6 @@ def fused_recurrent_kda(
         inplace_final_state=inplace_final_state,
         cu_seqlens=cu_seqlens,
         # ssm_state_indices=ssm_state_indices,
-        num_accepted_tokens=None,
         use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
     )
     return o, final_state

python/sglang/srt/layers/attention/mamba/mamba_state_scatter_triton.py

@@ -18,8 +18,6 @@ def _fused_mamba_state_scatter_with_mask_kernel(
     # Raw index arrays (before index_select)
     dst_indices_raw_ptr,  # [total_requests] - state_indices_tensor
     step_indices_raw_ptr,  # [total_requests] - accepted_steps or mamba_steps_to_track
-    # Total number of requests
-    total_requests,
     elem_per_entry: tl.constexpr,
     src_layer_stride,
     src_req_stride,
@@ -176,7 +174,6 @@ def fused_mamba_state_scatter_with_mask(
         dst,
         dst_indices_raw,
         step_indices_raw,
-        total_requests,
         elem_per_entry,
         src_layer_stride,
         src_req_stride,

python/sglang/srt/layers/attention/trtllm_mla_backend.py

@@ -59,7 +59,7 @@ def pad_draft_extend_query_kernel(
     q_ptr,  # Input query tensor [total_seq_len, num_heads, head_dim]
     padded_q_ptr,  # Output padded query tensor [batch_size, max_seq_len, num_heads, head_dim]
     seq_lens_q_ptr,  # Sequence lengths for each sequence [batch_size]
-    cumsum_ptr,  # Cumulative sum of accept lengths [batch_size + 1]
+    cumsum_ptr,  # Cumulative sum of sequence lengths [batch_size + 1]
     batch_size,
     max_seq_len,
     num_heads,
@@ -78,7 +78,7 @@ def pad_draft_extend_query_kernel(
     if batch_id >= batch_size:
         return
 
-    # Load accept length for this batch
+    # Load sequence length for this batch
     seq_len = tl.load(seq_lens_q_ptr + batch_id)
 
     if seq_pos >= seq_len: