[Spec Decode] Fix input triton kernel for eagle (vllm-project#15909)

Signed-off-by: xinyuxiao <[email protected]>

Author: ekagra-ranjan

vllm/v1/spec_decode/eagle.py

@@ -250,13 +250,12 @@ def prepare_input_kernel(
     num_tokens = end_pos - start_pos
 
     index_start = tl.load(cu_query_lens_ptr + pid)
-    indices = index_start + tl.arange(0, BLOCK_SIZE)
 
     num_blocks = tl.cdiv(num_tokens, BLOCK_SIZE)
     for i in tl.range(num_blocks):
         offset = i * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
         tl.store(
             out_ptr + start_pos + offset,
-            indices,
+            index_start + offset,
             mask=offset < num_tokens,
         )