Conversation
|
/rerun-test test_page_tail_indices.py test_dllm_fdfo_kv_reuse.py test_unified_radix_cache_kl_full.py test_streaming_session_swa.py |
|
Results for 🚀 🚀 🚀 |
|
/tag-and-rerun-ci |
|
/rerun-test test_page_tail_indices.py test_hisparse_allocator.py test_disaggregation_hisparse.py test_disaggregation_decode_radix_cache.py test/registered/disaggregation/test_disaggregation_basic.py |
|
Results for 🚀 🚀 🚀 🚀 |
# Conflicts: # python/sglang/srt/mem_cache/allocation_sizing.py
|
/rerun-test test_page_tail_indices.py test_streaming_session_swa.py test_unified_radix_cache_kl_full.py test/registered/disaggregation/test_disaggregation_basic.py test_disaggregation_hisparse.py |
|
Results for 🚀 🚀 🚀 🚀 |
|
/tag-and-rerun-ci |
The paged allocator hands out whole pages, but the row write only covers the requested tokens, so
req_to_token[:, seq_len : ceil(seq_len)]keeps the previous occupant's indices. Fill that tail (slots inside a page are consecutive) and widen the row headroom topage_size - 1for everypage_size > 1config, not just speculative ones.CI States
Latest PR Test (Base): ❌ Run #32304945912
Latest PR Test (Extra): ❌ Run #32304945576
Latest PR Test (AMD ROCm 10): ❌ Run #32304945813