minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix tensor mismatch error in sepc + topk > 1 + page_size > 1 (#14874)

Browse Source
This commit is contained in:
Ziming Huang
2025-12-15 01:10:58 +08:00
committed by GitHub
parent 47633c191f
commit 997ea57eaf
2 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/flashattention_backend.py
+3 -2
View File
@@ -2519,7 +2519,8 @@ def update_draft_decode_set_expand_metadata_with_page_size(
last_page_lens_broadcast = expanded_last_page_lens.unsqueeze(-1).expand( last_page_lens_broadcast = expanded_last_page_lens.unsqueeze(-1).expand(
-1, expand_page_table.shape[1] -1, expand_page_table.shape[1]
) )
expand_page_table -= last_page_lens_broadcast num_seqs = expand_page_table.shape[0]
expand_page_table -= last_page_lens_broadcast[:num_seqs]
expand_page_table = ( expand_page_table = (
expand_page_table[ expand_page_table[
:, strided_indices_expand[: (decode_length + page_size - 1) // page_size] :, strided_indices_expand[: (decode_length + page_size - 1) // page_size]
@@ -2527,4 +2528,4 @@ def update_draft_decode_set_expand_metadata_with_page_size(
// page_size // page_size
) )
max_seq_pages_expand = (decode_length + page_size - 1) // page_size max_seq_pages_expand = (decode_length + page_size - 1) // page_size
page_table[:, :max_seq_pages_expand].copy_(expand_page_table) page_table[:num_seqs, :max_seq_pages_expand].copy_(expand_page_table)
test/registered/spec/eagle/test_eagle_infer_a.py
+1
View File
@@ -228,6 +228,7 @@ class TestEAGLERadixCache(CustomTestCase):
# Chunked prefill & Page Size > 1 # Chunked prefill & Page Size > 1
{**self.BASE_CONFIG, "chunked_prefill_size": 64, "page_size": 4}, {**self.BASE_CONFIG, "chunked_prefill_size": 64, "page_size": 4},
{**self.BASE_CONFIG, "page_size": 4}, {**self.BASE_CONFIG, "page_size": 4},
{**self.BASE_CONFIG, "cuda_graph_bs": [5], "page_size": 4},
# Preferred by some kernels # Preferred by some kernels
{**self.BASE_CONFIG, "page_size": 64}, {**self.BASE_CONFIG, "page_size": 64},
# Disable CUDA Graph # Disable CUDA Graph