mirrors/sglang
0
0 Fork 0
mirror of https://gitee.com/trusted-list/sglang.git synced 2026-10-28 16:45:11 +08:00
Code Activity

Commit Graph

  • 7c98507040 Remove unnecessary comments in fp8_utils.py xpu/qwen3-fp8-w8a16-dense Alex Nails 2026-10-28 00:59:41 -07:00
  • 2d3ef161ff [Fix] [XPU] quantization: Honor weight_scale strides in W8A16 dequant, guard fused-op layouts Alex Nails 2026-10-28 00:51:00 -07:00
  • f313c035e8 [Perf] [XPU] quantization: Run block-FP8 linears weight-only (W8A16) on Xe2 Yihao Wang 2026-10-27 09:10:15 +00:00
  • 168154d20c [Fix] [XPU] benchmark: Run the fused MoE Triton tuner on XPU xpu/qwen3-moe-tuning-xpu Yihao Wang 2026-10-27 06:11:12 +00:00
  • 49657e2b5a [Perf] [XPU] qwen3: Fuse q/k RMSNorm and RoPE on XPU xpu/qwen3-qwen3-fused-qknorm-rope Yihao Wang 2026-10-27 02:26:52 +00:00
  • 520b855d09 [Perf] [XPU] distributed: Default oneCCL to its Arc all-reduce on Arc GPUs xpu/qwen3-ccl-arc-allreduce Yihao Wang 2026-10-27 02:26:09 +00:00
  • c536c44202 [Fix] [XPU] spec: Sample EAGLE verification on XPU instead of forcing argmax xpu/qwen3-eagle-target-only-sampling Yihao Wang 2026-10-27 05:44:49 +00:00
  • 6e61eca8ee [Fix] [XPU] spec: Give EAGLE drafts triton attention when intel_xpu lacks their head size xpu/qwen3-eagle-draft-attn-fallback Yihao Wang 2026-10-27 05:15:06 +00:00
  • e7fbf03f74 [Fix] [XPU] rope: Fall back to native RoPE for head sizes the XPU kernel lacks xpu/qwen3-rope-native-fallback Yihao Wang 2026-10-27 04:09:19 +00:00
  • 6aeefd6986 test: supply state tree slots in remaining MLX component fixtures cheng/refactor/unified-kv-tree-components Cheng Wan 2026-10-28 00:57:53 -07:00
  • 33d1b50773 Migrate the sharded eviction fixture to prefix-tree slots Cheng Wan 2026-10-27 23:22:21 -07:00
  • 74f6259256 Migrate the LLaDA image encoder to prefix-tree slots Cheng Wan 2026-10-27 22:38:23 -07:00
  • 4fe46104f9 Complete resolved geometry in the state cleanup test fixture Cheng Wan 2026-10-27 22:16:23 -07:00
  • ddfb3e24d1 unified-kv-tree-components(native-slot-families,non_mechanical_provable): Build the native tree core's drivers by slot family Cheng Wan 2026-10-27 15:05:23 -07:00
  • deb15daf6e unified-kv-tree-components(family-names,non_mechanical_provable): Name the prefix tree's slots by family in the cache and both tree cores Cheng Wan 2026-10-27 14:50:32 -07:00
  • 9cabb5e2d1 unified-kv-tree-components(tracking-slots,non_mechanical_provable): Take and give back a request's state tracking slots through the KV storage Cheng Wan 2026-10-27 12:56:09 -07:00
  • caa0e25caa unified-kv-tree-components(tree-slot-numbers,non_mechanical_provable): Carry each tree slot's window and checkpoint interval from the plan Cheng Wan 2026-10-27 12:33:24 -07:00
  • b30898206a unified-kv-tree-components(state-effects,non_mechanical_provable): Reach the prefix tree's recurrent state through the KV storage Cheng Wan 2026-10-27 08:05:36 -07:00
  • 349ec7bec4 unified-kv-tree-components(tree-slot-families,non_mechanical_provable): Build the prefix tree's components by slot family from the KV storage Cheng Wan 2026-10-27 05:37:54 -07:00
  • b4a89881f2 unified-kv-tree-components(tree-effects,non_mechanical_provable): Route the prefix tree's frees and window rebinds through the KV storage Cheng Wan 2026-10-27 02:31:19 -07:00
  • 91df012641 unified-kv-tree-components(advance-window,non_mechanical_provable): Retire a request's window rows through its KV storage Cheng Wan 2026-10-27 01:26:52 -07:00
  • 3db4e8572d unified-kv-tree-components(tree-slots,non_mechanical_provable): Build the prefix tree's components from the KV storage Cheng Wan 2026-10-27 01:23:06 -07:00
  • 1086d4ad14 fix: initialize FlashInfer test routing through its pool cheng/refactor/unified-kv-typed-views Cheng Wan 2026-10-28 00:24:41 -07:00
  • 8e6d38426d Migrate new speculative decoding fixtures to runner identifiers Cheng Wan 2026-10-27 23:22:10 -07:00
  • 2924c73b31 Preserve masked window planning through FlashInfer forward Cheng Wan 2026-10-27 22:48:46 -07:00
  • b334474081 Recognize MLA views when configuring HiCache storage Cheng Wan 2026-10-27 22:39:48 -07:00
  • 502fa981bb Mask out-of-window ancestors in speculative verification trees Cheng Wan 2026-10-27 22:02:29 -07:00
  • 99d00cfc6f Fix routed MLA writes to select the layer storage locations Cheng Wan 2026-10-27 22:02:13 -07:00
  • c029cdbf7c Fix fused draft HiCache validation after view binding Cheng Wan 2026-10-27 22:01:56 -07:00
  • b962a7d31e unified-kv-typed-views(id-modules,non_mechanical_provable): Keep a runner's KV id translation beside the id spaces Cheng Wan 2026-10-27 20:14:21 -07:00
  • 8315cb88ed unified-kv-typed-views(drop-spec-declarations,non_mechanical_provable): Build every speculative forward without per-backend combination lists Cheng Wan 2026-10-27 18:59:10 -07:00
  • 93143fb813 unified-kv-typed-views(trtllm-window-draft-extend,non_mechanical_provable): A trtllm_mha draft extend reads a window layer through the plan's window table Cheng Wan 2026-10-27 18:58:47 -07:00
  • a0bfe9b373 unified-kv-typed-views(flashinfer-window-read,non_mechanical_provable): the verify kit's FlashInfer DFLASH window is per query Cheng Wan 2026-10-27 18:43:51 -07:00
  • 5c968d90e6 unified-kv-typed-views(flashinfer-window-read,non_mechanical_provable): FlashInfer's kernel over a window read, checked against positions Cheng Wan 2026-10-27 18:18:32 -07:00
  • 0b5b89cdaa unified-kv-typed-views(flashinfer-window-read,non_mechanical_provable): one window read for every FlashInfer mode Cheng Wan 2026-10-27 18:16:12 -07:00
  • 26c900cd57 unified-kv-typed-views(callers-kv-layers,non_mechanical_provable): Reach per-layer KV through the layer router in the remaining callers Cheng Wan 2026-10-27 17:57:07 -07:00
  • 3ec1e82ac0 unified-kv-typed-views(other-backends-kv-layers,non_mechanical_provable): Read and write the remaining attention backends' per-layer KV through the layer router Cheng Wan 2026-10-27 17:57:06 -07:00
  • 6c20a42d0b unified-kv-typed-views(aiter-kv-layers,non_mechanical_provable): Read and write AITER's per-layer KV through the pool's layer router Cheng Wan 2026-10-27 17:57:05 -07:00
  • 4f157a97cc unified-kv-typed-views(models-mla-kv-layers,non_mechanical_provable): Read and write the remaining models' per-layer KV through the layer router Cheng Wan 2026-10-27 16:59:20 -07:00
  • c2d92ef8ac unified-kv-typed-views(mla-kv-layers,non_mechanical_provable): Read and write the MLA backends' per-layer KV through the layer router Cheng Wan 2026-10-27 16:59:20 -07:00
  • 5c02b627c0 unified-kv-typed-views(trtllm-torch-kv-layers,non_mechanical_provable): Read and write TRT-LLM MHA's and the torch backends' per-layer KV through the layer router Cheng Wan 2026-10-27 16:51:36 -07:00
  • 4d54fd3aae unified-kv-typed-views(models-kv-layers,non_mechanical_provable): Read and write models' per-layer KV through the active pool's layer router Cheng Wan 2026-10-27 16:51:36 -07:00
  • dc06fd3c88 unified-kv-typed-views(flashinfer-kv-layers,non_mechanical_provable): Read and write FlashInfer's per-layer KV through the runner's layer router Cheng Wan 2026-10-27 16:42:35 -07:00
  • 7adc3601e0 unified-kv-typed-views(fa3-kv-layers,non_mechanical_provable): Read and write FlashAttention's per-layer KV through the runner's layer router Cheng Wan 2026-10-27 16:42:35 -07:00
  • e427a2d815 unified-kv-typed-views(triton-kv-layers,non_mechanical_provable): Read and write Triton's per-layer KV through the runner's layer router Cheng Wan 2026-10-27 16:41:09 -07:00
  • af62163fe3 unified-kv-typed-views(kv-layer-access,non_mechanical_provable): Read and write a layer's KV through its router's KVLayer Cheng Wan 2026-10-27 16:40:56 -07:00
  • dbef5a6a2e unified-kv-typed-views(draft-views,non_mechanical_provable): Read a fused draft's KV and state through typed views of its host entries Cheng Wan 2026-10-27 12:35:17 -07:00
  • 3343cccbf5 unified-kv-typed-views(state-view,non_mechanical_provable): Read, clear and copy recurrent state through a typed view of the arena Cheng Wan 2026-10-27 12:32:49 -07:00
  • 81ebfe212d unified-kv-typed-views(mla-view,non_mechanical_provable): Read and store MLA latent rows through a typed view of the arena Cheng Wan 2026-10-27 12:30:32 -07:00
  • 1624edd242 unified-kv-typed-views(mha-view,non_mechanical_provable): Read and store MHA KV through a typed view of the arena Cheng Wan 2026-10-27 12:19:32 -07:00
  • a558f73489 style: format static KV storage transfer boundary cheng/refactor/unified-kv-storage-consumers Cheng Wan 2026-10-28 00:53:58 -07:00
  • 675d5486c3 unified-kv-static-transfer(pins-and-sizing,non_mechanical_provable): Register transfer pins and size HiCache alike for both placements Cheng Wan 2026-10-26 22:01:51 -07:00
  • c5a5fda900 unified-kv-static-transfer(static-pd-surface,non_mechanical_provable): Register static pools with PD peers through the storage interface Cheng Wan 2026-10-26 22:01:50 -07:00
  • 6a385b1f5c unified-kv-observers(beam-rows,non_mechanical_provable): Take and give back beam member rows and their ids through the KV storage Cheng Wan 2026-10-26 22:04:36 -07:00
  • 45bc0d6112 unified-kv-observers(usage,non_mechanical_provable): Read the pool observers' numbers from the KV storage's per-sub-pool usage Cheng Wan 2026-10-26 22:04:29 -07:00
  • a5d88ad44c unified-kv-lifecycle(lifecycle-through-storage,non_mechanical_provable): Take a request's rows and KV ids through the KV storage cheng/refactor/unified-kv-requests Cheng Wan 2026-10-26 19:58:59 -07:00
  • b4b37bbdcd unified-kv-requests(request-table,non_mechanical_provable): Hold every unified plan's request rows in one RequestTable Cheng Wan 2026-10-26 19:03:44 -07:00
  • c51c0b8c00 unified-kv-id-spaces(unit-ids,non_mechanical_provable): Keep a sub-pool's unit names in its own record cheng/refactor/unified-kv-arena-and-ids Cheng Wan 2026-10-26 18:33:59 -07:00
  • 4dc90e1090 unified-kv-id-spaces(bands,non_mechanical_provable): Place each unified sub-pool's units in a band of the arena Cheng Wan 2026-10-26 18:26:24 -07:00
  • 0258081e8c unified-kv-id-spaces(id-space-home,non_mechanical_provable): Define the id space in the KV allocation package Cheng Wan 2026-10-26 18:21:49 -07:00
  • ff6df073a4 unified-kv-memory-and-ids(storage-result,non_mechanical_provable): Return the unified KV storage from the builder Cheng Wan 2026-10-26 17:45:12 -07:00
  • e78544d4f4 unified-kv-memory-and-ids(arena,non_mechanical_provable): Hold the unified placement's bytes in a KVArena of bands Cheng Wan 2026-10-26 17:27:45 -07:00
  • 266e581a62 fix: use resolved decode graph configuration for speculative capabilities cheng/refactor/unified-kv-draft-views Cheng Wan 2026-10-28 00:52:54 -07:00
  • 400086a358 unified-kv-draft-views(drop-draft-backend-lists,non_mechanical_provable): Fuse a draft whatever attention backend reads it Cheng Wan 2026-10-26 16:53:49 -07:00
  • 5120120a9b unified-kv-draft-views(routed-draft,non_mechanical_provable): Route a fused draft's layers to its host sub-pools' views Cheng Wan 2026-10-26 16:53:20 -07:00
  • 568669aebe fix: retain one current-role transfer pin across PD switches cheng/refactor/unified-kv-pins Cheng Wan 2026-10-28 00:51:30 -07:00
  • 8db0f341d2 unified-kv-pins(host-ids-from-storage,non_mechanical_provable): Map host transfer ids through the KV storage instead of the device views Cheng Wan 2026-10-26 16:46:33 -07:00
  • 391da7416b unified-kv-pins(pin-sources,non_mechanical_provable): Keep unified pages in place through pin sources registered on the KV storage Cheng Wan 2026-10-26 16:46:08 -07:00
  • af9e5981a7 unified-kv-drop-facades(combination-ratchet,non_mechanical_provable): Fail a new unified pool class named for a combination of KV kinds cheng/refactor/unified-kv-drop-facades Cheng Wan 2026-10-26 16:06:51 -07:00
  • 1cc372255b unified-kv-drop-facades(routed-kv-cache,non_mechanical_provable): Route a unified pool's KV through one RoutedKVCache and delete the per-combination facades Cheng Wan 2026-10-26 16:05:12 -07:00
  • 506129af97 unified-kv-drop-facades(consumer-questions,non_mechanical_provable): Ask the allocator and the router what the unified KV facades answered Cheng Wan 2026-10-26 16:05:04 -07:00
  • 13042505d3 unified-kv-pins(hip-fused-window-write,non_mechanical_provable): Leave a unified window's KV writes to the pool on ROCm cheng/hot-fix/unified-kv-hip-fused-kv-write Cheng Wan 2026-10-26 16:35:12 -07:00
  • e1644d157b unified-kv-transfer(transfer-compare,non_mechanical_provable): Add a script comparing what PD peers and HiCache see of a unified pool cheng/refactor/unified-kv-transfer Cheng Wan 2026-10-26 14:14:12 -07:00
  • 5fc472243b unified-kv-transfer(pd-from-plan,non_mechanical_provable): Register a unified pool with PD peers from its plan Cheng Wan 2026-10-26 14:14:00 -07:00
  • 070d3844d2 unified-kv-transfer(hicache-from-plan,non_mechanical_provable): Build a unified pool's HiCache host stack from its plan Cheng Wan 2026-10-26 14:13:38 -07:00
  • 053f96c727 unified-kv-transfer(storage-names,non_mechanical_provable): Give each unified sub-pool its external names and a transfer surface Cheng Wan 2026-10-26 14:13:03 -07:00
  • c512039b81 unified-kv-transfer(tri-host-stack,non_mechanical_provable): Build the full+SWA+state HiCache stack like the other hybrid stacks cheng/hot-fix/unified-kv-tri-host-stack Cheng Wan 2026-10-26 13:35:29 -07:00
  • 2faf40c84a fix: bind EAGLE buffers through KV layer routes cheng/refactor/unified-kv-state-access Cheng Wan 2026-10-27 23:32:37 -07:00
  • 71c5f0b7e2 Update GDN graph fixtures for state views and identifier spaces Cheng Wan 2026-10-27 23:12:04 -07:00
  • 238b2fd8c7 Route GDN prefill graph state through the KV layer view Cheng Wan 2026-10-27 23:06:47 -07:00
  • abc20183b2 unified-kv-outside-branches(kind-facts,non_mechanical_provable): Read combined kind checks outside the pools as plan facts Cheng Wan 2026-10-26 12:27:02 -07:00
  • 3943b6df97 unified-kv-outside-branches(pool-predicates,non_mechanical_provable): Ask the KV pool's router what PD and HiCache dispatch on Cheng Wan 2026-10-26 12:26:51 -07:00
  • ca2831e222 unified-kv-state-access(state-built-by-builder,non_mechanical_provable): Build the unified state storage in the pool builder Cheng Wan 2026-10-26 12:03:09 -07:00
  • 2aadb602a5 unified-kv-state-access(state-acquire,non_mechanical_provable): Bind a request's row and state slots through one acquire Cheng Wan 2026-10-26 10:38:01 -07:00
  • 273f607416 unified-kv-state-access(state-readers,non_mechanical_provable): Reach recurrent state through the KV pool's router and the token allocator Cheng Wan 2026-10-26 10:11:27 -07:00
  • 0f0f1a76ca unified-kv-state-access(state-storage,non_mechanical_provable): Hold a table's recurrent state in one StateStorage Cheng Wan 2026-10-26 10:06:26 -07:00
  • c669c71e48 unified-kv-state-access(state-routes,non_mechanical_provable): Read recurrent state through the KV pool's layer router Cheng Wan 2026-10-26 09:53:10 -07:00
  • 83e2b5a85e unified-kv-state-access(clear-slots-no-sync,non_mechanical_provable): Clear unified state slots without a host synchronization Cheng Wan 2026-10-26 08:27:56 -07:00
  • 007434e24a unified-kv-state-access(state-attrs,non_mechanical_provable): Read state pool attributes without getattr Cheng Wan 2026-10-26 08:20:51 -07:00
  • 317f6eb73f unified-kv-state-access(state-id-space,non_mechanical_provable): Map state slot ids through the allocator's id space Cheng Wan 2026-10-26 08:05:55 -07:00
  • 6d0c45528f unified-kv-router(drop-chain-window-translate,non_mechanical_provable): Remove ChainAllocator.translate_loc_from_full_to_swa cheng/refactor/unified-kv-router Cheng Wan 2026-10-26 07:09:48 -07:00
  • f447a3f6b5 unified-kv-router(backend-routing,non_mechanical_provable): Route attention backends' window and layout checks through the layer router Cheng Wan 2026-10-26 07:09:26 -07:00
  • b71814774b unified-kv-router(router-and-views,non_mechanical_provable): Route each layer to the sub-pool view that holds its KV Cheng Wan 2026-10-26 07:08:53 -07:00
  • 9f0af30395 unified-kv-addressing(drop-target-allowlists,non_mechanical_provable): Stop gating the unified target's attention backend by speculative algorithm cheng/refactor/unified-kv-addressing Cheng Wan 2026-10-26 05:31:46 -07:00
  • 13d111c864 unified-kv-addressing(call-sites,non_mechanical_provable): Map window ids through id spaces at every call site Cheng Wan 2026-10-26 05:26:31 -07:00
  • 9429cd5324 unified-kv-addressing(id-spaces,non_mechanical_provable): Build the KV index translator from the pools' id spaces Cheng Wan 2026-10-26 05:26:18 -07:00
  • ed46981e6d unified-kv-request-lifecycle(chain-budget,non_mechanical_provable): Admit every unified chain through one prefill budget that owns its state-slot gate cheng/refactor/unified-kv-pricing-and-budget Cheng Wan 2026-10-26 05:11:00 -07:00
  • 3ca3103467 unified-kv-planner(snapshot-memo,non_mechanical_provable): Keep the pricing snapshot between decisions and price on resolved bands Cheng Wan 2026-10-26 05:20:31 -07:00
  • 6f60f12534 unified-kv-planner(pricing-property,non_mechanical_provable): Check that each capacity price is what the allocation then does Cheng Wan 2026-10-26 04:05:10 -07:00
  • e5afb50cd9 unified-kv-planner(snapshot-pricing,non_mechanical_provable): Price chain capacity on a host-only snapshot and take through one path Cheng Wan 2026-10-26 04:03:33 -07:00