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

[CI] Run test_unified_radix_cache_kl_dcp on cutedsl_mla with bf16 KV cache (#39553)

Browse Source
This commit is contained in:
Khoa Pham
2026-09-15 00:35:01 -07:00
committed by GitHub
parent 8a20062652
commit e4a6b090f4
1 changed files with 2 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py
+2 -4
View File
@@ -4,7 +4,7 @@ Under DCP the radix layer allocates widened logical indices while each rank's
buffers hold only its 1/dcp_size shard, so a missing translation makes cache buffers hold only its 1/dcp_size shard, so a missing translation makes cache
hits return another rank's KV. The KL cases catch that as a large divergence. hits return another rank's KV. The KL cases catch that as a large divergence.
Blackwell-only: the MLA DCP decode path needs ``tokenspeed_mla`` (SM100/12x). Blackwell-only: ``cutedsl_mla`` is the DCP-native MLA decode kernel on SM100.
""" """
import unittest import unittest
@@ -65,9 +65,7 @@ class TestUnifiedKimiLinearDcpHiCache(UnifiedRadixTreeTestMixin, CustomTestCase)
"--page-size", "--page-size",
str(PAGE_SIZE), str(PAGE_SIZE),
"--attention-backend", "--attention-backend",
"tokenspeed_mla", "cutedsl_mla",
"--kv-cache-dtype",
"fp8_e4m3",
"--dcp-comm-backend", "--dcp-comm-backend",
"a2a", "a2a",
"--dcp-replicate-q-proj", "--dcp-replicate-q-proj",