Repository navigation
Capture KV host transfer buffers - #43711
Merged
Merged
Conversation
ispobock
requested review from
Ying1123,
hanming-lu,
hnyls2002,
merrymercy and
xiezhq-hermann
as code owners
October 11, 2026 15:53
ispobock
added
run-ci-extra
ispobock
requested review from
alphabetc1,
huangtingwei9988 and
hzh0425
as code owners
October 11, 2026 15:53
Collaborator
Author
|
/rerun-test test/registered/unit/mem_cache/pool_host/test_kvcacheio_asymmetric.py test/registered/unit/mem_cache/test_asymmetric_mha_pool_host_unit.py test/registered/unit/mem_cache/test_hicache_dcp_host_pool.py test/registered/unit/mem_cache/test_hicache_host_register.py test/registered/unit/mem_cache/test_hicache_seaweedfs_storage.py test/registered/unit/mem_cache/test_hicache_staged_write_back_dispatch.py test/registered/unit/mem_cache/test_mem_pool_host.py test/registered/unit/mem_cache/test_mla_host_dedup_primitives.py |
Collaborator
Author
|
/rerun-test test/registered/hicache/test_hicache_variants.py test/registered/hicache/test_qwen35_hicache.py test/registered/hicache/test_hicache_spec_file_storage.py test/registered/e2e/hicache/test_hicache_unified_memory.py test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py test/registered/kernels/ops/kvcache/test_hicache_strided_device_rows.py test/registered/kernels/ops/kvcache/test_mxfp8_host_allocation.py |
Contributor
|
🚀 🚀 🚀 🚀 🚀 |
Contributor
|
🚀 🚀 |
3 of 5 tasks
reger-men
added a commit
to reger-men/sglang
that referenced
this pull request
Oct 12, 2026
pool_host/base.py: import conflict only, main dropped the KVCache import, kept take_contiguous_run. alloc() and _merge_release_slots() are unchanged on main. pool_host/mla.py merged without a conflict but the layer_first kernel reload still passed device_pool.kv_buffer[device_layer_id] to bulk_reload. After sgl-project#43711 that name only exists on the kernel_ascend path, so point bulk_reload at the captured self.device_kv_buffers[layer_id] like the jit and transfer_kv_per_layer_mla calls next to it (this also picks the right buffer for packed MTP draft layers). Signed-off-by: adlashab <adlashab@amd.com>
Liwansi
added a commit
to Liwansi/sglang_vans
that referenced
this pull request
Oct 13, 2026
…n_allgather * 'main' of https://github.com/sgl-project/sglang: (236 commits) [NPU] Stabilize HiCache and registered CI tests (sgl-project#43308) [Docs] Cookbook: refresh SemiAnalysis AgentX launches (GLM-5.2 B300, DeepSeek-V4.1 GB300) (sgl-project#43799) [Docs] Update GLM-5.3-Flash playground: B300 default, CP8, EP and PD (sgl-project#43802) [DSV4.1][PD][1/N] Support prefill node use cp parallel strategy (sgl-project#40323) [BCG][DSV4-Flash][CP] Support DeepSeek V4 prefill with interleave CP and breakable CUDA graphs (sgl-project#38603) [XPU] Support decode context parallelism (DCP) on Intel XPU (sgl-project#34355) Keep NIXL Mamba checkpoint keys separate by TP rank (sgl-project#43594) Capture KV host transfer buffers (sgl-project#43711) [CI] Move tests that do not need GB300 off the per-PR GB300 suite (sgl-project#43784) [HiCache] Preserve all MTP KV and recurrent state (sgl-project#40223) [HiCache] Support Mamba in buffer-only host memory mode (sgl-project#36345) [PD] Poison the deferred abort ack when a Mooncake send fails (sgl-project#43719) [Perf] Route BF16 mixed-batch decode tails through decode attention (sgl-project#43238) [Fix] Link cuda_ipc JIT against CUDA driver stubs dir (sgl-project#43764) docs: clarify prefill CP terminology and simplify wording (sgl-project#43787) [diffusion][xpu] FLUX.2: enable the bit-exact Triton SwiGLU on XPU (sgl-project#41940) [Fix] Adapt to transformers 5.18 / 5.19 changes (sgl-project#43688) [XPU CI] Optimize PR CI per-file timeouts and fail hanging tests fast (sgl-project#42048) [Deps] Bump FlashInfer to 0.7.1 (sgl-project#43494) docs: sync LMSYS SGLang blog cards (sgl-project#43462) ...
This was referenced Oct 14, 2026
Jianhong-Zhang
added a commit
to Jianhong-Zhang/sglang
that referenced
this pull request
Oct 21, 2026
* [diffusion] perf: pipeline MiniMax-H3's Ulysses exchange against cuDNN SDPA on Blackwell (sgl-project#43373) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] fix: fail MiniMax-H3 quality="high" closed on substituted DiT weights (sgl-project#34978) Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * [diffusion] Z-Image: autotune the shared transformer block at both token scales (sgl-project#35755) * [DSv4.1] Initialize the full-to-SWA mapping for request-window pools (sgl-project#43165) * [Dsv4.1] Add DeepSeek V4.1 DSPARK TP and DP EP accuracy coverage (sgl-project#43478) * [Refactor][TCPCG] Retire backend selection and configuration (3/9) (sgl-project#42287) * [CI] Remove the GLM-5.3-Flash PTX KDA prefill nightly test (sgl-project#43470) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * Clamp default decode CUDA graph max_bs to max_running_requests (sgl-project#43172) * [Model Loader] Add InstantTensor loader support (sgl-project#40453) * [CI] Fix Python 3.10 failure in prefill complete integration test (sgl-project#43505) * [diffusion] fix: keep a LoRA unmerged under merge_mode auto when merging would round its update away (sgl-project#43385) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [AMD] Resolving bug in per batch size graph capture due to multiple capture runs per batch size (sgl-project#39840) Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: HAI <hixiao@gmail.com> * [AMD][CI] Drop removed tp_rank/tp_size kwargs from small-M FP8 proj test (sgl-project#43318) * [AMD] Add an opt-in AITER CK blockscale GEMM for gfx942 block-FP8 linears (sgl-project#41071) Signed-off-by: siliangchen-amd <SiLiang.Chen@amd.com * [CI] Drop CPU tests with no bug-catching signal and run Rust workspace tests in their own workflow (sgl-project#43503) * [rust-server] preserve PD routing in chat and completions (sgl-project#42989) Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * [AMD][CI] Cap pyarrow below 26 for lmms-eval (sgl-project#43547) * [Rust] Upgrade sglang-radix-tree to PyO3 0.29 (sgl-project#42807) Co-authored-by: colleen-tml <296620403+colleen-tml@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [CI] Fix EAGLE3 CUDA parity prefill chunk alignment (sgl-project#43538) * [diffusion] perf: pipeline every USPAttention Ulysses path's exchange over the copy engine (sgl-project#43521) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] perf: run head_dim-128 QK-norm + RoPE two rows per warp for every RoPE layout (sgl-project#43508) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [Responses] Fix max_output_tokens overestimate on multimodal requests (sgl-project#35524) * [HiSparse] Avoid host synchronization in eager backup (sgl-project#41446) * [MemCache] refactor: require an anchored lock receipt for non-root releases (sgl-project#41452) * [HiCache] Add staged write-back for the page-unified KV cache layout (sgl-project#39606) * [kv-shard 3/4] Enable Control Plane C (sgl-project#40102) * [Docs] Document Qwen3-4B/8B H200 mixed-chunk latency tradeoffs (sgl-project#43551) * [Docs] Let the Qwen3 B300 recipe use the default attention backend (sgl-project#43554) * [HiCache] Keep protected cgroup page cache charged in host budgets (sgl-project#42788) Co-authored-by: Ke Bao <ispobaoke@gmail.com> * [Scheduler] Use committed length in spec mamba track boundary check (sgl-project#43029) Co-authored-by: Ke Bao <ispobaoke@gmail.com> Co-authored-by: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com> * [Refactor] Record what a bound stage path does instead of reading it back (sgl-project#43453) * [Fix] Keep residual rows and sums consistent across appends and pipeline handoffs (sgl-project#43454) * [Refactor] Make residual read and update capabilities explicit and checked (sgl-project#43455) * [Refactor] Decide input-scattered eligibility and stage rejections in one place (sgl-project#43456) * [Refactor] Bind a layer stack in one pass over each stage's data flow (sgl-project#43457) * [Refactor] Decide an exit's fixed facts when the stage is built (sgl-project#43458) * [Refactor] Declare pipeline neighbours from a shared declaration function (sgl-project#43459) * [Refactor] Declare every model's stages from a shared function and drop the meta-device neighbour build (sgl-project#43460) * [CI] Fix Kimi-Linear DCP PD test startup (sgl-project#43598) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * [Model] Keep gpt-oss router logits batch-invariant under deterministic inference (sgl-project#43624) * [Dsv4.1] Mega-mhc integration (sgl-project#42808) Co-authored-by: Yuwei An <yuwei.an@MacBook-Pro-yuweian.local> Co-authored-by: kpham-sgl <khoa.pham@radixark.ai> * [diffusion] fix: accept request-level profiling on the image API (sgl-project#43573) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] feat: declare regional compile regions for flux (sgl-project#43575) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] feat: bind each GPU worker to its GPU's NUMA node (sgl-project#43565) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] feat: warn instead of disabling on unvalidated models for BCG(sgl-project#43568) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] feat: keep the TP all-reduce inside the graph (sgl-project#43602) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] fix: keep exported buffers out of expandable segments for IPC all-to-all(sgl-project#43584) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] fix: skip QKV packing while layers are offloaded (sgl-project#43576) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] perf: add the dynamic LoRA delta inside the second GEMM (sgl-project#43501) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] CI: skip the LTX-2 RMSNorm-modulation guard test off the CUDA platform (sgl-project#43574) Co-authored-by: Mick <mickqian@radixark.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] fix: skip inductor's saved multi-kernel pick for torch.compile (Z-Image warm starts crash or hang) (sgl-project#43601) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] comfyui: fail early instead of silently dropping features in SGLD integrated mode (sgl-project#43495) Co-authored-by: OnePunchMonk <aggarwal.avaya@yahoo.com> * [Fix] Reserve remaining tokens for ignore_eos requests during prefill admission (sgl-project#43564) * [DSA] Support TileLang FP8 KV cache on CUDA (sgl-project#42957) Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> * fix: wait for HiCache load-back before hybrid MLA KV reads (sgl-project#38217) Co-authored-by: zhouzechun <zhouzechun@kuaishou.com> Co-authored-by: alphabetc1 <2508695655@qq.com> * [diffusion] fix: keep the overlay probe from stalling cached startup (sgl-project#43590) * [diffusion] fix: serve whole graphs after offload-during-compile (sgl-project#43631) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] Support LLaDA-Image BF16 generation and editing (sgl-project#37907) * [PD][HiCache] Fold the decode restore outcome into the MIN-reduced KV poll (sgl-project#43562) Co-authored-by: Kevin Flansburg <kflansburg@cloudflare.com> * [diffusion] perf: keep Qwen-Image-2.1 and Ulysses graphs whole (sgl-project#43586) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] perf: keep Cosmos3's fused kernels in the graph (sgl-project#43577) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] fix: Qwen-Image BCG prompt padding misses every graph and attends the pad (sgl-project#43632) Co-authored-by: Mick <mickqian@radixark.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] fix: replay Qwen-Image-2.1 BCG graphs across prompt lengths (sgl-project#43633) Co-authored-by: Mick <mickqian@radixark.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] perf: run the attention section eager under torch.compile for MiniMax-H3 (sgl-project#43634) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [1/N][DSv41] Enable DP attention with MegaMoE (sgl-project#43228) * [diffusion] comfyui: consolidate FluxAdapter correctness fixes (sgl-project#43490) Co-authored-by: niehen6174 <1639206518@qq.com> * [diffusion] comfyui: consolidate executor conditioning-cache fixes (sgl-project#43492) Co-authored-by: niehen6174 <1639206518@qq.com> * [Scheduler] Keep waiting-prefix LRU refresh consistent across TP ranks (sgl-project#43479) * [Perf] Fuse BF16 writes to native HND KV cache storage (sgl-project#43237) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [diffusion] Don't report video jobs completed after their outputs are deleted (sgl-project#43084) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> * [GDN] Fix CuTe DSL decode to read the [N, HV, V, K] state layout (sgl-project#43332) Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> * [KDA] Support Helion MTP verification and harden prefill, decode (sgl-project#42857) * [AMD]Qwen add Sparse attention to capturing hip graphs (sgl-project#43383) * [Spec] Allow Qwen3.5 backends to opt into packed target verification (sgl-project#43037) * [qwen 3.8 next] Fuse QSA prefill KV gathering and dtype conversion (sgl-project#41933) * [Scheduler] Requeue preempted requests when admission admits nothing (sgl-project#43298) * [Test] Do not hang scripted-runtime teardown when the server died (sgl-project#43285) * [KV canary] Expect sweep counters on the target runner only (sgl-project#43286) * [Scheduler] Hold back decode rows on a mixed step when they do not fit the KV pool (sgl-project#43290) * [Sampler] Sort only the max(top_k) leading columns in the torch sampler (sgl-project#43346) * [Scheduler] Keep a DP-converted decode batch as the running batch (sgl-project#43283) * [kvevents] fix missing topics in replayed kv events (sgl-project#43386) Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * [DeepSeek-V4.1] Use num_stages=1 for the hc_mix stats partial kernel (sgl-project#43070) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Yuan Luo <4908075+yuan-luo@users.noreply.github.com> * [VLM] fix: repair multimodal VMM lease cleanup and source-stream ordering (sgl-project#43475) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * [Diffusion] Add ModelOpt FP8 checkpoint quantization for Pi0.5 (sgl-project#43062) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * fix: default Inkling to fa4 on Hopper (sgl-project#35292) * [BugFix][DP] Prevent health checks from perturbing user request routing (sgl-project#35721) Signed-off-by: Phoenix3334 <Phoenix3334@users.noreply.github.com> * [Unified Radix] Invalidate lazy Mamba checkpoints before in-place tracking (sgl-project#43567) * [diffusion] perf: encode LTX video/audio through the shared bounded MP4 path (sgl-project#43656) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * [HiCache] Keep storage workers alive on backend exceptions (sgl-project#43550) * Build Mamba host pools from device buffer infos (sgl-project#43596) * Share MLA host capacity initialization (sgl-project#43595) * Preserve serialized page dtype in NIXL copy fallback (sgl-project#43593) * [diffusion] fix: select VSA-H3 for FastH3 when no transformer backend is given (sgl-project#43690) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [KV Events] Publish LoRA adapter names and namespaced block hashes (sgl-project#42455) Signed-off-by: Banruo Liu <lichangxinlivingshade@gmail.com> Co-authored-by: Shangming Cai <csmthu@gmail.com> * [DSpark] Run every markov head through one step interface (sgl-project#43620) * [DSpark] Add `MarkovSampleStep`, one op for a markov draft step's add and token pick (sgl-project#43621) * [DSpark] Draft the eager markov block through `MarkovSampleStep` (sgl-project#43622) * [DSpark] Run the same markov block in the draft graph tail; drop `SGLANG_DSPARK_FOLDED_SAMPLING` (sgl-project#43668) * [DSpark] Merge the draft TP sync sites into `dspark-draft` (sgl-project#43669) * [DSpark] Add `SGLANG_DSPARK_USE_GREEDY_DRAFT`, an argmax draft accepted as one-hot (sgl-project#43670) * [DSpark] Apply `SGLANG_SIMULATE_ACC_LEN` inside the accept instead of switching paths (sgl-project#43623) * [diffusion] fix: shard Qwen-Image adaLN modulation and text FFN under TP (sgl-project#43691) * [diffusion] perf: layerwise offload for the LTX-2.5 Gemma-4 text encoder (sgl-project#43683) * [diffusion] perf: split and batch LTX-2.3 one-stage guidance passes under CFG parallel (sgl-project#43684) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * [Fix] Initialize BF16 GEMM dispatch in the one-batch benchmark (sgl-project#43702) * Fix Mistral-native format detection when the Hub is unreachable (sgl-project#43446) * [MoE] Skip block-padding rows in the TMA down activation and quant (sgl-project#42212) * [MemCache] Split the SWA mapping move out of locked-full recovery as `RecoverEvictedSWA` (sgl-project#43721) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: Jialin <Jialin@users.noreply.github.com> * [Rust] Lock hyper 1.12.0 and want 0.3.2 in the rust workspace (sgl-project#43724) Co-authored-by: hanminglu <hanminglu@users.noreply.github.com> Co-authored-by: shodoco <shodoco@users.noreply.github.com> * [MemCache] Add shared component size and path APIs to the tree core interface (sgl-project#43722) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: Jialin <Jialin@users.noreply.github.com> * [CI] Fix TestQwenImageModulationCache after Qwen-Image modulation sharding (sgl-project#43747) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [Bugfix] Preserve available host memory for integrated MUSA devices (sgl-project#43398) Signed-off-by: Xucheng Zhou <aden1350@outlook.com> * [Tokenizer] Require fastokens>=0.3.2 and include it in the all extra (sgl-project#43104) * [AMD][V4.1][*/N] Fix DP attention + TP MoE decode NaN and support bounded replay (sgl-project#43077) Co-authored-by: wunhuang <wunhuang@amd.com> Co-authored-by: HAI <hixiao@gmail.com> * Support kv cache sharding for DSA indexer (sgl-project#40925) * [AMD][MORI] Add MORI EP V2 support (sgl-project#40666) * [PD] Retract transferred decode requests on a retract pause (sgl-project#43757) * [DSV4.1] Reduce eager Engram extend preparation (sgl-project#42159) Co-authored-by: Shangming Cai <csmthu@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [Test] Re-anchor the VLM flashinfer throughput floor (sgl-project#43762) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * [Fix] scheduler: stop latching batch_is_full when nothing can run (sgl-project#41872) Co-authored-by: hnyls2002 <lsyincs@gmail.com> Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com> * [RL] Return a failed weight update when the cache flush is refused (sgl-project#43758) * [CI] Skip eval retries in the PD transfer-failure test (sgl-project#43769) * docs: sync LMSYS SGLang blog cards (sgl-project#43462) Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com> * [Deps] Bump FlashInfer to 0.7.1 (sgl-project#43494) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * [XPU CI] Optimize PR CI per-file timeouts and fail hanging tests fast (sgl-project#42048) * [Fix] Adapt to transformers 5.18 / 5.19 changes (sgl-project#43688) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * [diffusion][xpu] FLUX.2: enable the bit-exact Triton SwiGLU on XPU (sgl-project#41940) Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * docs: clarify prefill CP terminology and simplify wording (sgl-project#43787) * [Fix] Link cuda_ipc JIT against CUDA driver stubs dir (sgl-project#43764) * [Perf] Route BF16 mixed-batch decode tails through decode attention (sgl-project#43238) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> * [PD] Poison the deferred abort ack when a Mooncake send fails (sgl-project#43719) Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> * [HiCache] Support Mamba in buffer-only host memory mode (sgl-project#36345) * [HiCache] Preserve all MTP KV and recurrent state (sgl-project#40223) Co-authored-by: alphabetc1 <2508695655@qq.com> Co-authored-by: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com> * [CI] Move tests that do not need GB300 off the per-PR GB300 suite (sgl-project#43784) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * Capture KV host transfer buffers (sgl-project#43711) * Keep NIXL Mamba checkpoint keys separate by TP rank (sgl-project#43594) * [XPU] Support decode context parallelism (DCP) on Intel XPU (sgl-project#34355) Co-authored-by: Siju Samuel <siju.samuel@intel.com> * [BCG][DSV4-Flash][CP] Support DeepSeek V4 prefill with interleave CP and breakable CUDA graphs (sgl-project#38603) * [DSV4.1][PD][1/N] Support prefill node use cp parallel strategy (sgl-project#40323) * [Docs] Update GLM-5.3-Flash playground: B300 default, CP8, EP and PD (sgl-project#43802) * [Docs] Cookbook: refresh SemiAnalysis AgentX launches (GLM-5.2 B300, DeepSeek-V4.1 GB300) (sgl-project#43799) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [NPU] Stabilize HiCache and registered CI tests (sgl-project#43308) * [sgl-router] Serve the Anthropic Messages API on /v1/messages (sgl-project#42316) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> * [Fix] Make tokenizer request-state initialization atomic (sgl-project#41911) Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com> * [XPU] Upgrade XPU backend to PyTorch 2.14 (sgl-project#42003) Co-authored-by: Ashwini Rathi <arathi@habana.ai> Co-authored-by: Y <ashwini.rathi@intel.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * [PD][Spec] Give a PD decode's first EAGLE draft step a proposal distribution (sgl-project#41783) Co-authored-by: Shangming Cai <csmthu@gmail.com> * [Fix] Adapt NPU Qwen-VL processor patch to transformers 5.19 resize() signature change (sgl-project#43309) * [NPU] Diffusion CI fixes (sgl-project#43210) * [diffusion] chore: bind worker CPUs to the GPU's NUMA node when memory policy is not permitted (sgl-project#43756) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [CI][Diffusion] Add caching for npu GT images in CI (sgl-project#37426) Co-authored-by: Артем Савкин <58187114+OrangeRedeng@users.noreply.github.com> * [QSA] Bound the decode indexer's work by the context length (sgl-project#43463) Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> * Prebuild fixup_zero_kv JIT module at TRTLLM-MLA backend init (sgl-project#37525) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * [Unified Cache] Abort requests whose external linker load fails instead of crashing the scheduler (sgl-project#40896) * [DSpark] Share draft params, logits dtype, confidence and the accept tail between graph and eager (sgl-project#43790) * [DFlash] Decide the Domino candidate pool in one place and apply simulated acceptance inside the accept (sgl-project#43791) * Fix autograd retention in CUDA graph capture tail hooks (sgl-project#43854) * [ROCm] Eliminate HiCache migration stalls and use exact-size host allocations (sgl-project#42553) Signed-off-by: JohnQinAMD <yanyuan.qin@amd.com> Co-authored-by: JohnQinAMD <yanyuan.qin@amd.com> * [AMD] Keep the FP4 varctx schedule test off AITER's 2048-row compile (sgl-project#41793) * [AMD] Add flydsl mxfp8 dense gemm kernel for dpsk v4.1 prefill (sgl-project#43803) * [AMD] Fix test_compact_attention_hip.py CI (sgl-project#43836) * [AMD] [Docker] Update MoRI to v1.2.4 (sgl-project#43834) * [Kernel] Expose bundled FlashMLA MegaAttention kernels and APIs (sgl-project#43918) * [DeepSeek-V4.1] Admit the NIXL transfer backend for DSpark PD (sgl-project#43734) * [AMD] Recalibrate test_quark_mxfp4 GSM8K thresholds to observed MI355X noise (sgl-project#41980) * [DSV4][CP] Write low-ratio page indices when CP prefill runs paged attention (sgl-project#43935) * [Build][CI] Drop Python 3.10: require Python 3.11+ and run CUDA CI on Python 3.12 (sgl-project#43021) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com> * [DeepSeek-V4.1] mHC: drop seam_open, narrow precomputed stats in take_rows (sgl-project#43833) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [DSv4.1] Reduce sparse-prefill workspace padding (sgl-project#43921) * [HiCache] Avoid full-store scans for uncached file prefix lookups (sgl-project#42077) Signed-off-by: apex-mochen <2756823972@qq.com> Co-authored-by: apex-mochen <265981728+apex-mochen@users.noreply.github.com> Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu> * [Fix] Preserve HTTP errors for non-streaming /v1/messages (sgl-project#43796) Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> * [DSv4.1] Temporarily disable B300 PD/CP accuracy test (sgl-project#43948) * [PD] Recompute retracted DeepSeek-V4 requests on prefill instead of crashing decode (sgl-project#43936) * [Fix] Account for captured inputs in DeepGEMM masked workspace budget (sgl-project#43563) * [AMD] Update v4 amd cookbook 1012 (sgl-project#43962) * [CI] Remove Mistral Small 4 e2e test (sgl-project#43973) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * Reuse Hugging Face LFM2-MoE configuration (sgl-project#43939) Co-authored-by: Codex <codex@openai.com> * fix(openai): honor tool_choice=none in DeepSeek encoders (sgl-project#43006) Co-authored-by: ai-jz <ai-jz@users.noreply.github.com> Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> * [diffusion] comfyui: run MiniMax H3, FastH3 and VDN in SGLD integrated mode (sgl-project#43504) * [diffusion] fix: keep MiniMax-H3 VAE fused norms off ROCm (sgl-project#43350) * [diffusion] fix: avoid redundant LoRA scaling while preserving exact numerics (sgl-project#43028) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * tokenizer perf optimization (sgl-project#39771) Co-authored-by: BBuf <1182563586@qq.com> Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> * [diffusion] perf: implement paint_use_remesh (decimate to 40k faces before UV unwrap) for Hunyuan3D (sgl-project#42770) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * [AMD][Kimi-K3] Use public FlyDSL APIs in KDA decode kernels (sgl-project#43968) Co-authored-by: wunhuang <wunhuang@amd.com> * [AMD] Fix Flux2 kv-gather abort on non-contiguous K/V (sgl-project#43793) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * [AMD] Fuse prefill index-Q RoPE fake-quant and fp4 quantize kernels in DSV4.1 (sgl-project#42911) * [AMD][DI][CI] dsv4pro-fp8 2p1d-ep16-mtp: raise prefill mem fraction for MORI EP v2 memory (sgl-project#43944) * [router] Fix power_of_two with --dp-aware (sgl-project#42027) * [ROCm][Spec] Route NGRAM tree reconstruction to the JIT kernel on HIP (sgl-project#43328) Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Thomas Wang <thomawan@amd.com> * [PD] Support heterogeneous TP for hybrid SWA KV transfer (sgl-project#40909) * [Diffusion] Speed up LTX-2 guidance and video output (sgl-project#43667) * [AMD][PD] enable mori transfer for PD decode DCP (sgl-project#37358) Signed-off-by: inkcherry <mingzhi.liu@amd.com> Co-authored-by: Duyi-Wang <duyi.wang@amd.com> Co-authored-by: HAI <hixiao@gmail.com> Co-authored-by: Shangming Cai <csmthu@gmail.com> * [Docs] Cookbook: refresh SemiAnalysis AgentX launches (DeepSeek-V4.1 GB200/GB300, GLM-5.2 FP8 B300) (sgl-project#43983) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [PD][NIXL] Fix staging buffer startup by accepting slot_layer_ids (sgl-project#43723) Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com> * [router] power_of_two: in-flight request load for --dp-aware decode (sgl-project#43483) * [Perf] Enable fast-math for JIT activations on Blackwell (sgl-project#43236) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [NVIDIA][MegaMoE] Fix the default spec decoding moe backend when the target model uses flashinfer megamoe (sgl-project#40772) Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com> * [HiCache][PD] Allow DSA index-K elision under HiCache and disaggregation (sgl-project#41769) Co-authored-by: Anish Shanbhag <ashanbhag@nvidia.com> * [CI] Use sgl-eval for GLM-5.3-Flash GSM8K tests (sgl-project#43975) Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * [FlashInfer] Add startup autotuning for MegaMoE (sgl-project#43789) * [Perf] Compact GDN causal-conv grids for skewed prefill batches (sgl-project#43570) * [AMD][CI] Run test_tracing.py in the extra tier, matching CUDA (sgl-project#43928) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Fix EVS for NemotronH_Omni_Reasoning_V3 (sgl-project#42933) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com> * fix(multimodal): convert grayscale tensors to RGB in PIL backend (sgl-project#42127) Co-authored-by: 周云昊 <zhouyunhao@zhouyunhaodeMacBook-Pro.local> * [diffusion] comfyui: run Qwen-Image 2.1 in SGLD integrated mode (sgl-project#43581) * [diffusion] feat: add configurable reference image resizing for MiniMax H3 Ref2VA (sgl-project#35839) Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> * Fix/compressed tensors nvfp4 hopper (sgl-project#40022) Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * [NPU] [DOC] update kimi k3 cookbook and supported features on NPU (sgl-project#44052) * [DSA] Q8KV8 FP8 Sparse Prefill on GLM-5.3-Flash with BF16 KV Cache (sgl-project#41790) Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> * [CI] Run test_pi05_modelopt_fp8_cuda in the kernel-unit lane (sgl-project#44055) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * fix(openai): omit absent tool_calls from chat responses (sgl-project#43027) Co-authored-by: ai-jz <ai-jz@users.noreply.github.com> Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> * Skip fully-masked KV tiles in sliding-window extend attention (sgl-project#39769) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(openai): keep chat stream created timestamp stable (sgl-project#43026) Co-authored-by: ai-jz <ai-jz@users.noreply.github.com> Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> * fix(openai): preserve zero cached tokens when cache reporting is enabled (sgl-project#43005) Co-authored-by: ai-jz <ai-jz@users.noreply.github.com> Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> * [AMD][DI][CI] glm52 2p1d-ep16: capture CUDA graph on decode (eager decode is too slow) (sgl-project#43977) * [AMD][DI][CI] kimik26 2p1d-ep16: capture CUDA graph on decode (eager decode is slow) (sgl-project#43985) * [perf] [multi-lora] remove empty CTAs from fused LoRA expand and add split-K to shrink (sgl-project#41951) Co-authored-by: Yanbin Jiang <jybsuper@gmail.com> * Rainj me/rust bench (sgl-project#43950) * [Observability] Report per-request storage prefetch and staging waits in meta_info (sgl-project#43946) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com> * [lora] Support LoRA for the NemotronH_Nano_VL_V2 model family (sgl-project#39798) Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Po-Han Huang <pohanh@nvidia.com> Co-authored-by: Yanbin Jiang <jybsuper@gmail.com> * [Rust] Refuse an unsupported multimodal model before the Rust listener binds (sgl-project#43925) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: shodoco <shodoco@users.noreply.github.com> * [Core] Keep request priority within the signed 64-bit range in radix caches (sgl-project#43920) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: qandrew <qandrew@users.noreply.github.com> * [DCP] Decode context parallelism for DSA (GLM-5.2 / GLM-5.3) with MTP (sgl-project#43735) Co-authored-by: vincentzed <207368749+vincentzed@users.noreply.github.com> * [HiCache] Block-group pool hits, caller-owned load destinations, allocation-target FULL eviction (sgl-project#43868) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com> * [Rust] Let an embedding package wrap the frontend HTTP app (sgl-project#43924) Co-authored-by: hanminglu <hanminglu@users.noreply.github.com> Co-authored-by: shodoco <shodoco@users.noreply.github.com> * [Core] Add request-level customized_info keys and mark Full prefill capture batches (sgl-project#43772) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: jiayisuse <jiayisuse@users.noreply.github.com> * [HiCache] Extract the buffer-mode prefetch and load-back lifecycle into pipeline operations (sgl-project#43870) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com> * [unified memory pool] Prefer SWA reclaim quotas with a HiCache storage backend or on PD decode (sgl-project#43919) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: ZYHowell <ZYHowell@users.noreply.github.com> * [Rust] Serve any `model` name on the OpenAI chat and completions routes (sgl-project#43923) Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: shodoco <shodoco@users.noreply.github.com> * docs: add MiniCPM-V 4.7 cookbook (sgl-project#43829) * [Feature] Add Aaronson-Gumbel text watermarking (sgl-project#37577) Co-authored-by: Nihar Patel <nihpat95@gmail.com> * [HiSparse] Add speculative decoding support with multi-step KV cache swapping (sgl-project#37771) * [gRPC] Move runtime.v1 bindings to tonic 0.14, report OpenAI status codes, return text on request (sgl-project#42673) Co-authored-by: Kan Wu <kan.wu@radixark.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [DeepSeek-V4.1] Run decoder SWA bounded replay inside the breakable prefill CUDA graph (sgl-project#43736) * fix(benchmark): request token IDs when counting OpenAI prompts (sgl-project#34511) Co-authored-by: Zijie Xia <zijie.xia@radixark.ai> * Mark FP32 redux asm volatile (sgl-project#44137) * [Score] Send stacked GPU embedding overrides through a CUDA IPC pool (sgl-project#39730) Co-authored-by: Sundara Raman Ramachandran <sundar24295@gmail.com> Co-authored-by: hnyls2002 <lsyincs@gmail.com> * [Feature] Return output top logprobs as flat base64 arrays on request (sgl-project#42890) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [DSV4.1] Support decoder SWA bounded replay with CUDA DP attention (sgl-project#43812) * GLM5.2 DSA Indexer Fallback (sgl-project#32947) Co-authored-by: Shu Wang <shuw@login-hecate03.hecate.clusters.nvidia.com> Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com> * [Fix] Work around a CPython fork deadlock in the CPU preprocess pool (sgl-project#44143) * [DSV4.1] Skip Mega mHC fusion on 0-row (DP-idle) batches (sgl-project#43993) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * [Chore] release: Cherry-pick [Disagg] Return prompt routed experts from the PD prefill server to release/v0.5.22 (sgl-project#44002) (sgl-project#44212) Co-authored-by: William Arnold <7565007+Aphoh@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com> * [Chore] release: Cherry-pick [CI] glm5.2: Temporarily disable PD HiSparse+MTP B300 test to release/v0.5.22 (sgl-project#44192) (sgl-project#44214) Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com> Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> * [XPU] Fix FP8 KV cache support for ModelOpt quantized models (e.g. Llama-3.3-70B-FP8) * [XPU] Add MXFP4 W4A16 MoE inference support for Intel Xe2 (GPT-OSS) Enable MXFP4 quantized inference on XPU using the native Xe2 SYCL fused_experts kernel (sgl-kernel-xpu PR#184). - quantization/__init__.py: register mxfp4 quantization method for XPU - quantization/fp4_utils.py: broaden flashinfer import exception to avoid crash on XPU where flashinfer is unavailable - quantization/mxfp4.py: - align intermediate size to 64 for XPU (create_weights) - skip CPU dequant on XPU; leave weights as packed uint8 in VRAM (process_weights_after_loading) - add XPU dispatch in apply(): call sgl_kernel.fused_experts with use_mxfp4_w4a16=True; convert UE8M0 uint8 scales to float32 via exp2(uint8 - 127); default gemm1_alpha=1.702 / gemm1_limit=7.0 to select activation_type=2 (ACT_SWIGLU_GPT_OSS, interleaved weight layout required by GPT-OSS MXFP4 checkpoints) - models/gpt_oss.py: replace hardcoded .cuda() with .to(get_device()) in weight loading to support XPU * [XPU] Improvement for fp4 and fp8 Optimize FP8 KV cache: lazy mirror buffer to avoid full cast per step Fix MXFP4 MoE: read activation/gemm1 params from model config instead of GPT-OSS hardcoded values * Remove WA for FP8 KV cache * Add analyze support for XPU * [Intel][XPU] Allow mamba extra_buffer radix cache strategy on XPU (#9) Co-authored-by: root <root@gnrsp-bmg3.iind.intel.com> --------- Signed-off-by: siliangchen-amd <SiLiang.Chen@amd.com Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> Signed-off-by: Phoenix3334 <Phoenix3334@users.noreply.github.com> Signed-off-by: Banruo Liu <lichangxinlivingshade@gmail.com> Signed-off-by: Xucheng Zhou <aden1350@outlook.com> Signed-off-by: JohnQinAMD <yanyuan.qin@amd.com> Signed-off-by: apex-mochen <2756823972@qq.com> Signed-off-by: inkcherry <mingzhi.liu@amd.com> Co-authored-by: Mick <mickjagger19@icloud.com> Co-authored-by: Mick Qian <mickqian@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Raymond Huang <103474329+Raymondlol@users.noreply.github.com> Co-authored-by: Frederik Gossen <frgossen@meta.com> Co-authored-by: YAMY <74099316+YAMY1234@users.noreply.github.com> Co-authored-by: Yuwei An <ayw.sirius19@gmail.com> Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com> Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai> Co-authored-by: cctry <csycfl@gmail.com> Co-authored-by: tao <yuanyitao@pku.edu.cn> Co-authored-by: Richard Wang <wangrichard08@gmail.com> Co-authored-by: mohbasit <mohbasit@amd.com> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: HAI <hixiao@gmail.com> Co-authored-by: chuyeh <chuyeh@amd.com> Co-authored-by: siliangchen-amd <SiLiang.Chen@amd.com> Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com> Co-authored-by: Sage <80211083+sagearc@users.noreply.github.com> Co-authored-by: Bingxu Chen <bingxche@amd.com> Co-authored-by: colleen-tml <colleend@thinkingmachines.ai> Co-authored-by: colleen-tml <296620403+colleen-tml@users.noreply.github.com> Co-authored-by: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com> Co-authored-by: hhy-seven <33918672+hhy-seven@users.noreply.github.com> Co-authored-by: huangtingwei <141888744+huangtingwei9988@users.noreply.github.com> Co-authored-by: Shunkangz <182541032+Shunkangz@users.noreply.github.com> Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> Co-authored-by: Ke Bao <ispobaoke@gmail.com> Co-authored-by: Khoa Pham <khoa.pham@radixark.ai> Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com> Co-authored-by: Yuwei An <yuwei.an@MacBook-Pro-yuweian.local> Co-authored-by: Mick <mickqian@radixark.ai> Co-authored-by: WenhaoZhang <42087078+niehen6174@users.noreply.github.com> Co-authored-by: OnePunchMonk <aggarwal.avaya@yahoo.com> Co-authored-by: Jzz1943 <jiangzhongze.jzz@ant-intl.com> Co-authored-by: LoCinder <61399298+LoCinder@users.noreply.github.com> Co-authored-by: zhouzechun <zhouzechun@kuaishou.com> Co-authored-by: alphabetc1 <2508695655@qq.com> Co-authored-by: Shangming Cai <csmthu@gmail.com> Co-authored-by: Kevin Flansburg <kflansburg@cloudflare.com> Co-authored-by: Avaya Aggarwal <119044997+OnePunchMonk@users.noreply.github.com> Co-authored-by: niehen6174 <1639206518@qq.com> Co-authored-by: Chunan Zeng <zcnrex@gmail.com> Co-authored-by: SunskyXH <SunskyXH@gmail.com> Co-authored-by: Nikolai <nprotaso@amd.com> Co-authored-by: paulzhang-tm <paulzhang@thinkingmachines.ai> Co-authored-by: Qiaolin Yu <liin1211@outlook.com> Co-authored-by: jiaqiang-dot-liu <jiaqiang.liu@amd.com> Co-authored-by: Ziyi Xu <76582120+DarkSharpness@users.noreply.github.com> Co-authored-by: Yuan Luo <4908075+yuan-luo@users.noreply.github.com> Co-authored-by: trebladev <54441939+trebladev@users.noreply.github.com> Co-authored-by: FeiHong Guo <1187883824@qq.com> Co-authored-by: Banruo Liu <lichangxinlivingshade@gmail.com> Co-authored-by: Alison Shao <54658187+alisonshao@users.noreply.github.com> Co-authored-by: WangHemeng <47976301+wanghemeng@users.noreply.github.com> Co-authored-by: metamergebot <metamergebot@gmail.com> Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com> Co-authored-by: Jialin <Jialin@users.noreply.github.com> Co-authored-by: hanminglu <hanminglu@users.noreply.github.com> Co-authored-by: shodoco <shodoco@users.noreply.github.com> Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com> Co-authored-by: adenzhou1350 <aden1350@outlook.com> Co-authored-by: weireweire <weiliangl@nvidia.com> Co-authored-by: kk <43161300+kkHuang-amd@users.noreply.github.com> Co-authored-by: wunhuang <wunhuang@amd.com> Co-authored-by: ChengYM <Yueming.Cheng@amd.com> Co-authored-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com> Co-authored-by: hnyls2002 <lsyincs@gmail.com> Co-authored-by: sglang-bot <sglangbot@gmail.com> Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com> Co-authored-by: ashwini rathi <ashwini.rathi@intel.com> Co-authored-by: Yihao Wang <42559837+AgainstEntropy@users.noreply.github.com> Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com> Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> Co-authored-by: Anupa Sajikumar <anupa.sajikumar@intel.com> Co-authored-by: Siju Samuel <siju.samuel@intel.com> Co-authored-by: BingjiaWang <wangbingjia.wbj@ant-intl.com> Co-authored-by: Jinyan Yi <yijinyan1@huawei.com> Co-authored-by: ybyang <10629930+whybeyoung@users.noreply.github.com> Co-authored-by: C <3121985758@qq.com> Co-authored-by: Pavan Sivaram Girijala <pavan.sivaram.girijala@intel.com> Co-authored-by: Ashwini Rathi <arathi@habana.ai> Co-authored-by: Sasha Sidorov <asidorov566@gmail.com> Co-authored-by: ming_wang <68357922+sigama-w@users.noreply.github.com> Co-authored-by: Артем Савкин <58187114+OrangeRedeng@users.noreply.github.com> Co-authored-by: Alexandr <117110413+Allor-maker@users.noreply.github.com> Co-authored-by: Yuan Luo <yuan.luo@hotmail.com> Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> Co-authored-by: Sam Shleifer <sshleifer@gmail.com> Co-authored-by: ix-48 <90045377+ix-48@users.noreply.github.com> Co-authored-by: Zhang, Jiejing <280536569+jiejingzhangamd@users.noreply.github.com> Co-authored-by: JohnQinAMD <yanyuan.qin@amd.com> Co-authored-by: Michael <13900043+michaelzhang-ai@users.noreply.github.com> Co-authored-by: Thomas Wang <thomawan@amd.com> Co-authored-by: Duyi-Wang <duyi.wang@amd.com> Co-authored-by: Yizheng Jiao <jyizheng@gmail.com> Co-authored-by: dingdangmao <2756823972@qq.com> Co-authored-by: apex-mochen <265981728+apex-mochen@users.noreply.github.com> Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu> Co-authored-by: Vegetog <110553275+Vegetog@users.noreply.github.com> Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> Co-authored-by: Piotr Mazurek <piotr635@gmail.com> Co-authored-by: Codex <codex@openai.com> Co-authored-by: ai-jz <156989844+ai-jz@users.noreply.github.com> Co-authored-by: ai-jz <ai-jz@users.noreply.github.com> Co-authored-by: Bingoo <33573610+BingooYang@users.noreply.github.com> Co-authored-by: jacky.cheng <yichiche@amd.com> Co-authored-by: Xinyi Song <xinyisong0111@gmail.com> Co-authored-by: Zhaoyi Li <36555117+Lzy17@users.noreply.github.com> Co-authored-by: AMD-yanfeiwang <yanfei.wang@amd.com> Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com> Co-authored-by: inkcherry <mingzhi.liu@amd.com> Co-authored-by: Michal Shalev <mshalev@nvidia.com> Co-authored-by: AMD-yanfeiwang <yanfwang@amd.com> Co-authored-by: Shu Wang <shuw@nvidia.com> Co-authored-by: Anish Shanbhag <ashanbhag@nvidia.com> Co-authored-by: Ziang Li <ziangli@umich.edu> Co-authored-by: elvischenv <219235043+elvischenv@users.noreply.github.com> Co-authored-by: Yunhao Zhou <131600287+yyh-sjtu@users.noreply.github.com> Co-authored-by: 周云昊 <zhouyunhao@zhouyunhaodeMacBook-Pro.local> Co-authored-by: LittleNyima <62497818+LittleNyima@users.noreply.github.com> Co-authored-by: jlqibm <87678480+jlqibm@users.noreply.github.com> Co-authored-by: amote-i <49533125+amote-i@users.noreply.github.com> Co-authored-by: Ho-Ren (Jack) Chuang <horenchuang@bytedance.com> Co-authored-by: Kailash Ranganathan <kailash@modal.com> Co-authored-by: Yanbin Jiang <jybsuper@gmail.com> Co-authored-by: Rain Jiang <96632942+rainj-me@users.noreply.github.com> Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com> Co-authored-by: Po-Han Huang <pohanh@nvidia.com> Co-authored-by: qandrew <qandrew@users.noreply.github.com> Co-authored-by: vincentzed <207368749+vincentzed@users.noreply.github.com> Co-authored-by: jiayisuse <jiayisuse@users.noreply.github.com> Co-authored-by: ZYHowell <ZYHowell@users.noreply.github.com> Co-authored-by: Nihar Patel <nihpat95@gmail.com> Co-authored-by: Kan Wu <wukanustc@gmail.com> Co-authored-by: Kan Wu <kan.wu@radixark.ai> Co-authored-by: Jiaqi Li <ljq.lijiaqi@outlook.com> Co-authored-by: Zijie Xia <zijie.xia@radixark.ai> Co-authored-by: Andrew Briand <atb8888@comcast.net> Co-authored-by: Chuanrui Zhu <47195419+chuanrui@users.noreply.github.com> Co-authored-by: Sundara Raman Ramachandran <sundar24295@gmail.com> Co-authored-by: HEJIAN SANG <sanghj0923@gmail.com> Co-authored-by: Shu Wang <shuw@login-hecate03.hecate.clusters.nvidia.com> Co-authored-by: shyeh25 <206795756+shyeh25@users.noreply.github.com> Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com> Co-authored-by: William Arnold <7565007+Aphoh@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Co-authored-by: Spurthi Lokeshappa <spurthi.lokeshappa@intel.com> Co-authored-by: root <root@gnrsp-bmg3.iind.intel.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
Capture MHA/MLA transfer buffers and pointer tables during host initialization. This removes repeated target/draft buffer selection from transfer callbacks and makes token-size calculation independent of initialization side effects.
Validation: 122 CPU tests and 34 subtests passed.
CI States
Latest PR Test (Base): ✅ Run #38153176669
Latest PR Test (Extra): ❌ Run #38153176492
Latest PR Test (AMD ROCm 10): ❌ Run #38153176661