MOVZX pushed to patches at iHore/vLLM 2026-10-18 16:04:41 +07:00
c5ab42bc0f Merge branch 'main' into patches
1b2a666391 [Doc][Recipe] Fix for redhat fp8 xeon support issue (#62245)
83c9a547bf [CI][Release] Fix release wheel builds: stop excluding tracked recipes/ from Docker context (#62221)
b510857d37 [Bugfix][Build] Build fused GDN MTP decode for SM80-SM90 with CUDA < 13 (#60896)
7c9042c05a [Bugfix][MRV2] Fix speculative warmup deadlock with pipeline parallelism (#62129)
Compare 36 commits »
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-18 15:38:53 +07:00
b1fdbde895 build: Rewrite artifacts/ when it is missing, and stop failing re-runs
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-18 06:29:20 +07:00
fe3cd9247b build: Reuse the extensions for a Python-only patch series
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-18 06:23:58 +07:00
50dabc9989 ple: drop the deferred state on a prefill-only step (avoid crashing)
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-18 02:05:23 +07:00
d9ca3bc4f2 feat: Set concurrency and the CUDA graph list from .env
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-18 01:49:30 +07:00
fab0eb9fc7 feat: Offload the vision tower to host RAM
MOVZX pushed to patches at iHore/vLLM 2026-10-17 15:22:31 +07:00
e0fe5df731 Merge branch 'main' into patches
08a65e287e [Bugfix][Kernel] hadacore: don't load past the input on a partial last warp (#61944)
ad2688ab30 [Bugfix] Validate draft-token capacity in forward benchmark (#61809)
5be4266f5c [Bugfix][Frontend] Preserve Cohere and Anthropic streaming errors (#61947)
90d6d871d6 [Attention] Add CuTe DSL backend selection for MLA context prefill (#62069)
Compare 50 commits »
MOVZX pushed to patches at iHore/vLLM 2026-10-16 17:35:49 +07:00
be95edb7f6 Merge branch 'main' into patches
07f3857fb9 [XPU][Linear][MXFP4] enable torch as mxfp4 linear backend on xpu (#51965)
fa75d3456d [Feature][PCP] Add DeepSeek-V4.1-Flash support (#59857)
074c83a09a [CI][Intel] Skip ROCm-only test_topk_gating_enabled_by_cdna on XPU (#61932)
e365792997 [Dependency] Upgrade TokenSpeed MLA to 2.0.3 (#61689)
Compare 65 commits »
MOVZX pushed to patches at iHore/vLLM 2026-10-16 16:08:13 +07:00
8dc54f35dd [Offloader] Stream every multimodal tower parameter to host RAM
MOVZX pushed to patches at iHore/vLLM 2026-10-16 14:32:27 +07:00
18c06ed406 [Offloader] Stream multimodal towers from host RAM with bulk-copy
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-16 13:01:46 +07:00
3a0a5d1510 feat: Set the engine stats interval to 1 s
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-16 12:40:58 +07:00
c4d88acb4d feat: Accept any model name in a request
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-16 01:53:27 +07:00
af239c4229 docs: Document the KV scale sidecar and the model variables
b89e85e99f feat: Mount the KV scale sidecar from any host path
1281bf9c03 fix: Load the FP8 KV scales from the shipped file name
Compare 3 commits »
MOVZX pushed to patches at iHore/vLLM 2026-10-15 22:56:40 +07:00
5d6754ff6b fix: keep the refusal-direction projection dynamo-safe
0076adc78b [Qwen3.5] Project the refusal direction out of the residual stream at runtime
76df77c04e Merge branch 'main' into patches
9cede34a67 [Bugfix] Create cuBLAS handles before memory profiling to avoid OOM (#61044)
252be28c18 [Frontend] Replace return_mm_kwargs with an mm_payload enum on render requests (#61427)
Compare 369 commits »
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-15 14:52:44 +07:00
d6062059b4 docs: Add the Indonesian README and sync the guide with the code
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-15 14:40:13 +07:00
27f50debe0 feat: Add benchmark wrapper script
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-15 14:27:41 +07:00
d365f4586f docs: Add power-limit benchmark results
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-15 13:37:43 +07:00
5cc9ee2505 chore: Reset vLLM logging interval back to default
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-15 13:36:34 +07:00
7e80cc67ee feat: Add pretty vLLM logging scripts
MOVZX pushed to main at iHore/vLLM-4x-RTX-3090 2026-10-15 13:33:44 +07:00
ae48511258 feat: Add pretty vLLM logging scripts