From e4a6b090f45f9eaece80c6abfbdb45ccd80b3627 Mon Sep 17 00:00:00 2001 From: Khoa Pham Date: Tue, 15 Sep 2026 00:35:01 -0700 Subject: [PATCH] [CI] Run test_unified_radix_cache_kl_dcp on cutedsl_mla with bf16 KV cache (#39553) --- .../unified_radix_tree/test_unified_radix_cache_kl_dcp.py | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py b/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py index 9a2b05e02..1d72c8612 100644 --- a/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py +++ b/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py @@ -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 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 @@ -65,9 +65,7 @@ class TestUnifiedKimiLinearDcpHiCache(UnifiedRadixTreeTestMixin, CustomTestCase) "--page-size", str(PAGE_SIZE), "--attention-backend", - "tokenspeed_mla", - "--kv-cache-dtype", - "fp8_e4m3", + "cutedsl_mla", "--dcp-comm-backend", "a2a", "--dcp-replicate-q-proj",