[Spec] Remove dead padded_static_len and stale SGLANG_ENABLE_SPEC_V2 references (#30998)

python/sglang/srt/batch_overlap/two_batch_overlap.py

@@ -732,7 +732,6 @@ class TboForwardBatchPreparer:
             "is_prefill_only",
             "spec_algorithm",
             "capture_hidden_mode",
-            "padded_static_len",
             "split_index",  # for split prefill
             "orig_seq_lens",  # only used by qwen-1m, thus not care
             "return_pooled_hidden_states",

python/sglang/srt/layers/logits_processor.py

@@ -230,8 +230,6 @@ class LogitsMetadata:
     global_num_tokens_for_logprob_gpu: Optional[torch.Tensor] = None
     # The gather mode for DP attention
     dp_padding_mode: Optional[DpPaddingMode] = None
-    # for padding
-    padded_static_len: int = -1
 
     # Whether this batch is prefill-only (no token generation needed)
     is_prefill_only: bool = False
@@ -279,7 +277,6 @@ class LogitsMetadata:
             top_logprobs_nums=forward_batch.top_logprobs_nums,
             token_ids_logprobs=forward_batch.token_ids_logprobs,
             extend_input_logprob_token_ids_gpu=forward_batch.extend_input_logprob_token_ids_gpu,
-            padded_static_len=forward_batch.padded_static_len,
             is_prefill_only=forward_batch.is_prefill_only,
             global_num_tokens_gpu=forward_batch.global_num_tokens_gpu,
             dp_local_start_pos=forward_batch.dp_local_start_pos,
@@ -527,21 +524,7 @@ class LogitsProcessor(nn.Module):
             and not logits_metadata.extend_return_logprob
         ):
             # Prefill without input logprobs.
-            if logits_metadata.padded_static_len < 0:
-                last_index = torch.cumsum(logits_metadata.extend_seq_lens, dim=0) - 1
-            else:
-                # If padding_static length is 5 and extended_seq_lens is [2, 3],
-                # then our batch looks like [t00, t01, p, p, p, t10, t11, t12, p, p]
-                # and this retrieves t01 and t12, which are the valid last tokens
-                idx = torch.arange(
-                    len(logits_metadata.extend_seq_lens),
-                    device=logits_metadata.extend_seq_lens.device,
-                )
-                last_index = (
-                    idx * logits_metadata.padded_static_len
-                    + logits_metadata.extend_seq_lens
-                    - 1
-                )
+            last_index = torch.cumsum(logits_metadata.extend_seq_lens, dim=0) - 1
             pruned_states = hidden_states[last_index]
             if hidden_states_before_norm is not None:
                 pruned_states_before_norm = hidden_states_before_norm[last_index]

python/sglang/srt/model_executor/forward_batch_info.py

@@ -502,9 +502,6 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
     dp_local_num_tokens: Optional[torch.Tensor] = None  # cached info at runtime
     global_dp_buffer_len: Optional[int] = None
 
-    # For padding
-    padded_static_len: int = -1  # -1 if not padded
-
     # For Qwen2-VL
     mrope_positions: torch.Tensor = None
 

python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py

@@ -130,7 +130,6 @@ class EAGLEDraftExtendCudaGraphRunner(DecodeCudaGraphRunner):
         self.capture_hidden_mode = CaptureHiddenMode.LAST
 
         self.capture_bs, _ = get_batch_sizes_to_capture(model_runner)
-        self.padded_static_len = -1
 
         # Size cuda-graph buffers by num_draft_tokens (full tree width), not
         # num_steps + 1, or topk > 1 draft-extend overflows them.
@@ -395,7 +394,6 @@ class EAGLEDraftExtendCudaGraphRunner(DecodeCudaGraphRunner):
             spec_algorithm=self.model_runner.spec_algorithm,
             spec_info=spec_info,
             capture_hidden_mode=CaptureHiddenMode.LAST,
-            padded_static_len=self.padded_static_len,
         )
 
         if self.buffers.dsa_seed_topk_capture is not None:

python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py

@@ -156,7 +156,6 @@ class MultiLayerEagleDraftExtendCudaGraphRunner(DecodeCudaGraphRunner):
         self.capture_hidden_mode = CaptureHiddenMode.FULL
 
         self.capture_bs, _ = get_batch_sizes_to_capture(model_runner)
-        self.padded_static_len = -1
 
         # Fixed window: every step extends each request by the same number of
         # tokens, which lets all steps share one buffer set.
@@ -301,7 +300,6 @@ class MultiLayerEagleDraftExtendCudaGraphRunner(DecodeCudaGraphRunner):
             capture_hidden_mode=capture_mode,
             extend_seq_lens=extend_seq_lens,
             extend_seq_lens_cpu=extend_seq_lens_cpu,
-            padded_static_len=self.padded_static_len,
             extend_start_loc=extend_start_loc,
             extend_num_tokens=self.num_tokens_per_req * bs,
             num_token_non_padded_cpu=self.num_tokens_per_req * bs,

python/sglang/test/kits/attention_unittest/runner_modes/split_op_runner.py

@@ -229,7 +229,6 @@ def _make_static_forward_batch(raw_batch, static_num_tokens: int, device: str):
         input_ids=input_ids,
         positions=positions,
         out_cache_loc=out_cache_loc,
-        padded_static_len=static_num_tokens,
         num_token_non_padded_cpu=raw_num_tokens,
     )