TokenSpeed Files

What's Changed

• Use CUDA routing gemm for LongCat by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/1
• Update README by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/2
• Add PyTorch index-url for rocm.txt requirements by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/3
• [AMD]Avoid Exceeding LDS Limit by @knwng in https://github.com/lightseekorg/tokenspeed/pull/4
• Skip FA4 registration on B300 (SM 10.3) by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/5
• Update README with new intro and performance diagram by @yubofredwang in https://github.com/lightseekorg/tokenspeed/pull/9
• fix: Kill server by PID in agentic_bench.sh by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/17
• fix: Avoid large FP32 init for skipped MLA ref checks by @huangzhilin-hzl in https://github.com/lightseekorg/tokenspeed/pull/12
• Support EvalScope box table score parsing by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/19
• decoder_layer: Python 3.10 compat (replace PEP 695 generics with TypeVar/Generic) by @andyluo7 in https://github.com/lightseekorg/tokenspeed/pull/23
• feat(PD): support mamba cache by @XucSh in https://github.com/lightseekorg/tokenspeed/pull/14
• platform: hardcode MI300/MI350 LDS size; probe /sys/class/kfd for xGMI topology by @andyluo7 in https://github.com/lightseekorg/tokenspeed/pull/25
• perf(K2.5): fix slow tokenize and add warmup for agentic bench by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/31
• perf(K2.5): Tune num_tokens for router_gemm and fused_a_gemm by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/33
• Add Kimi K2.5 agentic perf CI task by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/29
• Fix agentic perf CI triggers by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/34
• [AMD] Add eval ci and fix kv store addr on mi355 by @borontion in https://github.com/lightseekorg/tokenspeed/pull/36
• [AMD]Use Iris Communication Ops by Default by @knwng in https://github.com/lightseekorg/tokenspeed/pull/38
• perf: default TF32 on for FP32 cuBLAS (router GEMMs) by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/35
• fix: mamba prefix cache by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/15
• Tune speculative decoding defaults by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/40
• feat(deepseek-v4): add mega_moe and compressed KV perf path by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/30
• Fix GPT-OSS MXFP4 scale dtype by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/42
• perf: optimize qwen3.5 runtime prepare cost by @minedec in https://github.com/lightseekorg/tokenspeed/pull/32
• [AMD]Upgrade triton_kernels Package by @knwng in https://github.com/lightseekorg/tokenspeed/pull/37
• fix(scheduler): keep host_node_ref alive across forward FSM states by @roycho96 in https://github.com/lightseekorg/tokenspeed/pull/41
• MiniMax-M2 H100 FP8 optimizations by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/10
• perf(eviction): O(k log N) eviction via persistent LRU set by @qywu in https://github.com/lightseekorg/tokenspeed/pull/18
• Move benchmark scripts under runtime tests by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/44
• perf: avoid GPU sync in extend metadata to restore overlap for trtllm mha by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/45
• [metrics] 1/N: rename collectors and clean up dead code by @yubofredwang in https://github.com/lightseekorg/tokenspeed/pull/7
• Add workflow to update tokenspeed-kernel FlashInfer by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/46
• Fix FlashInfer update workflow branch push by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/48
• Update tokenspeed-kernel FlashInfer to 0.6.11 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/49
• Suppress FlashAttention JIT cache debug logs by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/50
• Default speculative draft quantization to unquantized by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/52
• Default prefill token budgets to 8k by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/53
• Align speculative config token steps by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/54
• Use FlashInfer wheel index for JIT cache by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/55
• Update tokenspeed-fa4 dependency by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/56
• eval: add qwen3.5 aime25 eval by @minedec in https://github.com/lightseekorg/tokenspeed/pull/43
• Remove runtime cutlass DSL dependency by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/59
• Move nvtx dependency to tokenspeed-kernel by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/60
• Update agent dependency guidance by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/61
• Suppress noisy CUTLASS DSL warnings by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/62
• Suppress TensorRT-LLM INFO startup logs by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/63
• Test MTP speculative config CLI parity by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/64
• Remove default eval prefill args by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/65
• test: cover tensor parallel aliases by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/66
• ci: use tp alias in eval configs by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/67
• Use packaged tokenspeed-triton-kernels dependency by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/68
• Use packaged tokenspeed-iris dependency by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/70
• Remove FlashInfer JIT cache requirement by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/71
• Enable FA4 prefill head dim 256 by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/73
• eval: temp set aime25 threshold to 0.86 by @minedec in https://github.com/lightseekorg/tokenspeed/pull/75
• perf(K2.5): fuse embeds and hidden norm in MLA eagle3 by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/78
• Fix tokenspeed proton import by @borontion in https://github.com/lightseekorg/tokenspeed/pull/79
• feat(deepseek-v4): add scheduler-managed sliding-window cache groups by @SimonCqk in https://github.com/lightseekorg/tokenspeed/pull/51
• fix(qwen3): plumb tensor-parallel info through MLP layers by @qywu in https://github.com/lightseekorg/tokenspeed/pull/80
• fix: runtime_stats.py -> runtime_states.py by @raayandhar in https://github.com/lightseekorg/tokenspeed/pull/82
• Refine third-party attribution notices by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/84
• feat: support num_heads = 64 for TokenSpeed MLA decode by @dishengbin in https://github.com/lightseekorg/tokenspeed/pull/85
• Add KV cache events to scheduler by @Aphoh in https://github.com/lightseekorg/tokenspeed/pull/8
• Refresh Python install paths to expose packages from subprocess pip by @borontion in https://github.com/lightseekorg/tokenspeed/pull/91
• Integrate attn backend with attn kernel selection by @borontion in https://github.com/lightseekorg/tokenspeed/pull/90
• Add --model-path alias for --model by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/92
• Fix KV cache event parent hash performance by @Aphoh in https://github.com/lightseekorg/tokenspeed/pull/87
• perf(K2.5): disable embedding TP and skip concat for EAGEL3 drafter loop by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/94
• feat(deepseek-v4): harden attention fast paths by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/93
• perf: optimize mamba prefix cache performance by @minedec in https://github.com/lightseekorg/tokenspeed/pull/77
• feat(cli): ts serve — smg gateway + gRPC engine by @lite-keke in https://github.com/lightseekorg/tokenspeed/pull/97
• ci: install smg deps from whl index by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/99
• ci: pin smg install deps by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/100
• chore: remove stale server and parsers by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/98
• Update SMG package pins by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/101
• fix(ci): fix reasoning parser for ci by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/102
• Package tokenspeed-mla binary objects by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/104
• Resolve platform-specific MLA binary objects by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/106
• Add tokenspeed-mla version update workflow by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/107
• Update tokenspeed-mla to 0.1.2 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/108
• Add tokenspeed-kernel MLA update workflow by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/109
• Update tokenspeed-kernel MLA to 0.1.2 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/110
• Test tokenspeed-mla binary prefill by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/111
• Use raw links for tokenspeed-mla images by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/112
• Default SMG serve port to 8000 by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/114
• Default SMG reasoning parser to none by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/115
• Remove redundant eval reasoning parser none by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/116
• Clean up stale server startup references by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/117
• test: add tokenspeed mla tests by @dishengbin in https://github.com/lightseekorg/tokenspeed/pull/118
• [fix] Correct gfx950 LDS size to 160 KB by @benenzhu in https://github.com/lightseekorg/tokenspeed/pull/119
• fix: support disabling KV prefix cache while keeping Mamba state cache functional by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/76
• fix to use max_num_pages for spec-decode topk page_table buffers by @roycho96 in https://github.com/lightseekorg/tokenspeed/pull/125
• perf(K2.5): enable AR-Norm fusion and fused FP8 decode for MLA Eagle3 by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/124
• perf: optimize flashinfer sampling backend by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/105
• perf(K2.5): Optimize lm_head by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/126
• feat(cli): print TokenSpeed banner on ts serve startup by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/127
• ci(cli): run CLI unit tests in ut-runtime-1gpu by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/129
• feat(cli): accept positional model arg in ts serve by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/128
• feat(cli): disable smg circuit breaker and retries by default by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/130
• Refine third-party attribution notices by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/131
• Fix sampling backend default on non-NVIDIA platforms by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/135
• feat(cli): default smg gateway log level to warn by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/136
• chore: drop legacy agentic_benchmark tree by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/137
• chore(ci): drop --stream and generation-config from gpt-oss gpqa eval by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/139
• fix: unset default max new tokens by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/141
• [AMD] Fix sampling backend from smg by @borontion in https://github.com/lightseekorg/tokenspeed/pull/143
• deps: switch SMG gateway packages to tokenspeed-smg by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/144
• feat(xgrammar): bring back structural tags of xgrammar to support json with reasoning by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/138
• feat(deepseek-v4): support mixed prefill/decode batches by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/122
• chore: cleanup noqa by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/147
• fix(cli): prewarm HF tokenizer cache before launching smg gateway by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/148
• perf(K2.5): optimize small kernels in EAGEL3 drafter loop by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/142
• fix(bench): add stream sock_read + per-request timeouts to unblock gather by @jasl in https://github.com/lightseekorg/tokenspeed/pull/134
• feat(models): add Qwen2 / Qwen2.5 dense LM support by @qywu in https://github.com/lightseekorg/tokenspeed/pull/39
• perf: limit grid for per-layer h2d copy by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/150
• fix(cli): pin smg prometheus port to 8413 by default by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/151
• ci(pipeline): pkill smg::router in stale-process cleanup by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/152
• fix(spec-decode): reject speculative_eagle_topk > 1 at startup by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/155
• chore: use StreamFork in DSv4 by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/157
• fix(server-args): restrict --speculative-eagle-topk to 1 by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/158
• perf: default turn on tokenizer cache. by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/160
• docs: update parser launch examples by @mmangkad in https://github.com/lightseekorg/tokenspeed/pull/163
• fix(runtime): widen BF16 CUTLASS MoE autotune range by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/161
• [AMD]Support a-fp8-w-mxfp4 gpt-oss-120b model by @knwng in https://github.com/lightseekorg/tokenspeed/pull/154
• refactor: forward mode consolidation by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/164
• fix(cache): emit AllBlocksCleared on reset only, not every match_prefix by @roycho96 in https://github.com/lightseekorg/tokenspeed/pull/166
• perf(cache): keep page-id tensors on GPU through free_group_end by @roycho96 in https://github.com/lightseekorg/tokenspeed/pull/167
• perf(sampling): fuse min-p renorm in flashinfer full backend by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/156
• feat(deepseek-v4): add persistent topk path by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/172
• perf: add gluon fp16 prefill kernel by @borontion in https://github.com/lightseekorg/tokenspeed/pull/165
• fix(deepseek-v4): corrected profiling to estimate cache capacity. by @SimonCqk in https://github.com/lightseekorg/tokenspeed/pull/173
• test: use mi350 platform to mean both mi350x/mi355x by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/175
• fix(server-args): bump tp4 gpu_memory_utilization to 0.95 by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/177
• perf(qwen3): cut H100 decode kernel time -8% with fused stride-aware kernels by @qywu in https://github.com/lightseekorg/tokenspeed/pull/81
• Perf[Qwen3.5]: eliminate Mamba intermediate state memcpy in MTP target-verify by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/159
• feat(mla): default prefill backend to binary, bump to 0.1.3 by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/180
• chore: remove duplicate LayerDoneCounte by @wtsung in https://github.com/lightseekorg/tokenspeed/pull/185
• feat(qwen3): add Qwen3 MoE causal LM support by @qywu in https://github.com/lightseekorg/tokenspeed/pull/181
• ci: add workflow to bump tokenspeed-smg* pyproject pins by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/186
• Update tokenspeed-smg* pins to 1.4.1.post20260519 / 0.4.7.post20260519 / 0.5.3.post20260519 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/187
• chore(cli): drop _overwrite_sampling_backend smg workaround by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/188
• fix scheduler req pool move assignment leak by @curious-whq in https://github.com/lightseekorg/tokenspeed/pull/182
• Fix Qwen3 FP8 MoE activation scale layout by @qywu in https://github.com/lightseekorg/tokenspeed/pull/189
• test(inline-detokenizer): accept hardware-variant reference outputs via multi-ref ROUGE-L by @qywu in https://github.com/lightseekorg/tokenspeed/pull/190
• feat: use split prefill for prefix cache in mha backend by @borontion in https://github.com/lightseekorg/tokenspeed/pull/178
• feat(deepseek-v4): overlap routed and shared MoE experts by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/192
• feat: add DeepSeek V4 parser support by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/193
• perf(K2.5): optimize top_k_renorm_prob + top_p_renorm_prob by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/184
• Tokenspeed MLA prefill kernel perf optimization by @dishengbin in https://github.com/lightseekorg/tokenspeed/pull/197
• chore(mla): default prefill backend back to cutedsl by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/201
• Update tokenspeed-kernel MLA to 0.1.3 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/202
• feat(deepseek-v4): device-side implement prefix caching for v4 hybrid cache. by @SimonCqk in https://github.com/lightseekorg/tokenspeed/pull/146
• refactor(deepseek-v4): clean up helper and kernel paths by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/203
• perf: add gluon fp16 decode kernel by @borontion in https://github.com/lightseekorg/tokenspeed/pull/191
• fix: fix the bug that kv cache pool size is too small. by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/194
• Avoid clearing FlashInfer cache in PR CI by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/209
• perf(qwen3.5): fuse q/k GemmaRMSNorm into one triton launch by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/196
• chore(ci): route agentic perf to b300-4gpu by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/211
• feat(kernel): add rope kernel api by @borontion in https://github.com/lightseekorg/tokenspeed/pull/208
• fix(deepseek-v4): refine cache sizing and shared expert comm by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/213
• perf: TokenSpeed MLA decode kernel optimization for num_heads=16 by @dishengbin in https://github.com/lightseekorg/tokenspeed/pull/212
• perf(K2.5): integrate cutedsl argmax by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/210
• perf(K2.5): support deterministic topk and remove maybe_broadcast by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/206
• Update tokenspeed-mla to 0.1.4 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/214
• Update tokenspeed-kernel MLA to 0.1.4 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/215
• Revert "Update tokenspeed-kernel MLA to 0.1.4 (#215)" by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/218
• NFC: Move MoE reference kernels to numerics by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/219
• NFC: Move NVFP4 SwiGLU GEMM wrapper to cute_dsl by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/220
• feat(kernel): add quantization kernel api by @borontion in https://github.com/lightseekorg/tokenspeed/pull/216
• test(kernel): lock public API kernel selection by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/221
• Revert "perf: TokenSpeed MLA decode kernel optimization for num_heads=16 (#212)" by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/222
• ci: test in-tree tokenspeed-mla when its sources change by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/223
• feat: mixed prefill-decode batching (part 3: compatibility with speculative decoding for MLA models) by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/205
• perf(qwen3.5): fuse attn_output_gate sigmoid+mul + drop gate reshape copy by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/198
• fix(deepseek-v4): enable DeepSeek-V4 unit tests on CI by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/224
• Update tokenspeed-mla to 0.1.5 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/225
• Update tokenspeed-kernel MLA to 0.1.5 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/226
• Add qwen3.5-397b-a17b nvfp4 perf CI task by @minedec in https://github.com/lightseekorg/tokenspeed/pull/195
• Perf[Qwen3.5]: some kernel fuse optimizations. by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/228
• fix(hf_transformers_utils): centralize architecture resolution; None-safe by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/231
• ci(pr-test): pass HF_TOKEN to unit-test jobs by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/233
• feat(ci): add optional priority field to task yaml by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/234
• perf[Qwen3.5]: fuse small kernels in MoE block. by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/235
• perf(qwen3.5): reduce prefill memcpy sync and mamba update overhead by @minedec in https://github.com/lightseekorg/tokenspeed/pull/229
• feat(ci): print perf-reference comparison table with % by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/237
• feat(kvstore): support mamba l2 cache transfers by @XucSh in https://github.com/lightseekorg/tokenspeed/pull/162
• fix(env): add missing runtime packages to PACKAGE_LIST by @yurekami in https://github.com/lightseekorg/tokenspeed/pull/240
• refactor(deepseek-v4): clean up attention metadata and cache helpers by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/242
• chore(qwen3.5): adjust qwen perf tps threshold by @minedec in https://github.com/lightseekorg/tokenspeed/pull/241
• feat(runtime): support multimodal VLM by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/236
• ci(perf-qwen3.5-agentic): route to gb200-4gpu-perf runner by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/245
• fix(kernel): drop Triton vendored CUDA include path from nvcc -I by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/247
• ci(perf): add Qwen3.5-NVFP4 agentic perf on b200-8gpu by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/250
• fix(logits): avoid nan in fused softcap by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/183
• chore(smg): bump deps to post20260525 + use passthrough parser by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/252
• feat(moe): allow Fp8FlashinferCutlassBackend on ep_size=1 when explicitly forced by @qywu in https://github.com/lightseekorg/tokenspeed/pull/244
• Fix: fix a hang bug in tp server. by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/248
• fix(deepseek-v4): shard attn_sink for tensor-parallel > 1 by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/254
• fix(deepseek-v4): correct admission over-credit for state-family paged-cache groups by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/249
• refactor: clean up triton kernel moe backend by @borontion in https://github.com/lightseekorg/tokenspeed/pull/267
• ci(eval): use --repeats 3 on gpt-oss-120b gpqa-diamond to suppress noise by @qywu in https://github.com/lightseekorg/tokenspeed/pull/278
• fix: missing causal mask for attn extend kernel by @borontion in https://github.com/lightseekorg/tokenspeed/pull/271
• fix(kvstore): decouple write and execute stream by @XucSh in https://github.com/lightseekorg/tokenspeed/pull/279
• perf(sampling): fuse logits fp32 cast to argmax or softmax by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/262
• fix(vlm): substitute mm pad ids in drafter before draft embed_tokens by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/265
• Update tokenspeed-smg* pins to 1.4.1.post20260527 / 0.4.8.post20260527 / 0.5.3.post20260527 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/283
• Use operand format signatures for kernel selection by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/230
• Make dtype-role filtering explicit by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/289
• perf: fuse kv write for decode only by @borontion in https://github.com/lightseekorg/tokenspeed/pull/281
• doc: Add news on README.md by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/292
• fix(vlm): clamp gpu_memory_utilization conditionally and add --language-model-only by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/266
• ci(eval): add Kimi-K2.5-NVFP4 ocr_bench task by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/253
• refactor(deepseek-v4): native deep_gemm FP8 GEMM + snapshot fix by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/288
• ci(perf): add 1m perf bench for qwen3.5 by @minedec in https://github.com/lightseekorg/tokenspeed/pull/264
• perf(Spec Decode): skip dead-position compute in draft catch-up step(decode) by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/217
• ci(eval): set kimi-k2.5 ocr-bench score_threshold to 0.9 by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/299
• ci(perf): add qwen3.5 agentic perf ci bs16 case by @minedec in https://github.com/lightseekorg/tokenspeed/pull/257
• refactor: drop legacy attn backends by @borontion in https://github.com/lightseekorg/tokenspeed/pull/307
• refactor: update forward metadata in mha backend by @borontion in https://github.com/lightseekorg/tokenspeed/pull/303
• perf: optimize logits allgather and parallelize eagle3 input projection by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/295
• Clean up stale GPU CI process groups by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/316
• skills: Add a skill to bisect triton release for regressions by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/322
• Document tokenspeed-kernel backend guidance by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/323
• deps(kernel): pin triton to 3.7.10.post20260505 by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/320
• Document dependency optionality guidance by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/324
• deps(kernel): update triton to 3.7.10.post20260531 by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/325
• Load upstream-produced inputs only after gridDepSync by @wangbo981016 in https://github.com/lightseekorg/tokenspeed/pull/327
• test: extend agentic bench by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/312
• fix: check mixed batch in mha backend based on forward mode by @borontion in https://github.com/lightseekorg/tokenspeed/pull/334
• feat(entrypoints): add HTTP server sidecar alongside smg gateway by @qywu in https://github.com/lightseekorg/tokenspeed/pull/308
• fix(deepseek-v4): close MTP acceptance gap by @Xiangyi1996 in https://github.com/lightseekorg/tokenspeed/pull/207
• fix(dp): fix qwen 3.5 data parallel bug. by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/309
• perf: Optimize inter-iteration small op by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/291
• chore: use -O3 -use_fast_math for tokenspeed_kernel compilation by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/285
• fix(PD): fix PD speculative bootstrap input seeding by @XucSh in https://github.com/lightseekorg/tokenspeed/pull/286
• Fix(spec decode): catch up trim bug by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/335
• perf(kernel): optimize mha kernel for sliding window case by @borontion in https://github.com/lightseekorg/tokenspeed/pull/336
• Defer perf reference failure reporting by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/340
• perf(deepseek-v4): decode attention optimizations by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/339
• perf: add Gluon MoE kernels for GPT-OSS by @knwng in https://github.com/lightseekorg/tokenspeed/pull/314
• feat: reduce DeepSeek V4 prefix state snapshots with replay reuse by @SimonCqk in https://github.com/lightseekorg/tokenspeed/pull/329
• Update tokenspeed-kernel FlashInfer to 0.6.11.post3 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/347
• [Perf] Optimizes loads in gfx950 fp16 decode kernel by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/313
• fix: NAN error in dp case cause of dummy input/data by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/341
• Drop nonexistent cuda_${SM} extra from TokenSpeed install by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/350
• fix(deepseek-v4): defer mega-MoE warmup and fix MoE TP weight loading by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/356
• perf(gdn): use flashinfer sm100 prefill kernel for Qwen3.5 GDN by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/331
• [Perf] Parallelize split-k reduce in mha_decode_fp16_gfx950 kernel by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/353
• chore(logits): drop prefill gather_ids fallback in LogitsProcessor by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/352
• fix(cute_dsl): skip CuTe DSL argmax kernel on H20 GPUs by @botieking98 in https://github.com/lightseekorg/tokenspeed/pull/311
• Update tokenspeed-smg* pins to 1.4.1.post20260607 / 0.4.8.post20260607 / 0.5.3.post20260607 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/371
• feat(engine): pause/resume scheduler control API by @HJSang in https://github.com/lightseekorg/tokenspeed/pull/346
• perf(deepseek-v4): decode kernel fusion and routing optimization by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/375
• Split vendor kernel packages by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/367
• Revert "Split vendor kernel packages" by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/378
• fix(cudagraph): set ctx.global_bs at capture for draft MoE all-gather by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/369
• feat(deepseek-v4): support MTP prefix cache reuse by @SimonCqk in https://github.com/lightseekorg/tokenspeed/pull/361
• ci(eval): add DeepSeek V4-Flash eval CI tasks by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/344
• [AMD] moe: fused single-kernel Gluon routing fast path for small-M decode (gfx950) by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/363
• fix: cpu buffer racing for sampling. by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/376
• Remove mini load balancer by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/384
• perf: TokenSpeed MLA decode kernel optimization for num_heads=16 by @dishengbin in https://github.com/lightseekorg/tokenspeed/pull/255
• Update tokenspeed-mla to 0.1.6 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/385
• Update tokenspeed-kernel MLA to 0.1.6 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/386
• Update tilelang to 0.1.11 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/387
• fix(ci): relax mm25 determinism guards to tolerate the -use_fast_math near-tie by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/388
• Add PD Qwen3.5 HTTP worker by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/389
• fix(scheduler): untrack TreeNode from mamba_leaves_ before prune frees it by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/357
• feat(gdn): replace gdn use h_track path with flashinfer prefill gdn by @minedec in https://github.com/lightseekorg/tokenspeed/pull/377
• Add tokenspeed mooncake dependency by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/391
• Add AMD Gluon sampling argmax by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/379
• Support DP sampling for spec decode by @yubofredwang in https://github.com/lightseekorg/tokenspeed/pull/232
• Retry ROCm pip installs by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/396
• fix(mamba): gate per-draft-token pool budget on speculative_algorithm by @rucnyz in https://github.com/lightseekorg/tokenspeed/pull/397
• feat(pd): heterogeneous tp support by @XucSh in https://github.com/lightseekorg/tokenspeed/pull/345
• ci(qwen3.5): add qwen3.5 397b pd ci (1p1d) by @minedec in https://github.com/lightseekorg/tokenspeed/pull/400
• Limit apache-tvm-ffi to 0.1.11 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/401
• Cap apache-tvm-ffi in CUDA requirements by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/402
• ci: add gpt-oss perf ci on mi350 by @borontion in https://github.com/lightseekorg/tokenspeed/pull/394
• feat: add mla backend and kernels by @borontion in https://github.com/lightseekorg/tokenspeed/pull/373
• perf(gdn): fused QKV split Triton kernel for GDN prefill by @elwhyjay in https://github.com/lightseekorg/tokenspeed/pull/380
• Update tokenspeed-deepgemm to 2.5.0.post20260604 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/410
• [CI] Use new label for gfx950 benchmark jobs by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/408
• ci: switch AMD 2GPU runner label by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/414
• Add small-M Gluon warp-decode MoE path for GPT-OSS by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/403
• Split AMD Gluon kernels package by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/404
• Fix ROCm CI AMD kernel install by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/417
• refactor(spec-decode): simplify Llama Eagle3 attention path for [#217] (1/3) by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/390
• fix(moe): fix divide zero bug and guard nvfp4 flashinfer forward on empty input by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/415
• perf(gdn): remove gdn prefill unnecessary h_state copy by @minedec in https://github.com/lightseekorg/tokenspeed/pull/409
• Add sm103a to default CUDA archs by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/420
• fix(kernel): Fix gluon MoE GEMM numerical bug by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/421
• perf(deepseek-v4): pre-compile deep_gemm JIT kernels at startup by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/398
• ci(gpt-oss): replace throughput to output throughput by @borontion in https://github.com/lightseekorg/tokenspeed/pull/422
• Update README status by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/424
• Sync env package list with dependencies by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/425
• Add Qwen3.5 PD AIME25 eval CI by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/426
• feat(kernel): introduce unified moe kernel api by @borontion in https://github.com/lightseekorg/tokenspeed/pull/374
• perf(deepseek-v4): dense deep_gemm warmup M-sweep + fp8_einsum coverage by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/427
• Update tokenspeed-smg dependency pins by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/428
• Add missing tokenspeed copyright headers by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/430
• Remove unused Python runtime helpers by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/431
• perf(kernel): optimize mha sliding window prefill kernel by @borontion in https://github.com/lightseekorg/tokenspeed/pull/432
• [AMD] warp_decode improvements by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/423
• ci: show AMD GPU information by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/435
• fix: limit internal activation dtype trait for moe plan by @borontion in https://github.com/lightseekorg/tokenspeed/pull/436
• ci: update AMD GPU runner labels by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/438
• fix(runtime bug): MoE collective sizing and aux hidden capture under heterogeneous TP by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/419
• perf(qwen3.5): use nvfp4_gemm_swiglu_nvfp4_quant in shared experts by @minedec in https://github.com/lightseekorg/tokenspeed/pull/433
• Upgrade transformers to 5.12.0 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/440
• perf(kernel): cap all-layer KV writeback grid to avoid starving concurrent compute by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/439
• Upgrade xgrammar to 0.2.2 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/441
• Relax openai dependency lower bound by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/442
• Remove unused tilelang dependency by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/443
• feat(moe): add trtllm mxint4 (INT4 W4A16 group-32) MoE path for Kimi-K2.x by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/444
• fix(cudagraph): seed warmup seq_lens to q_len to avoid NAN by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/446
• feat(engine): Sleep / Wake Up API (release/resume_memory_occupation) by @HJSang in https://github.com/lightseekorg/tokenspeed/pull/393
• chore(license): use LightSeek header on sleep/wake files by @HJSang in https://github.com/lightseekorg/tokenspeed/pull/449
• fix pd decode TP4->DP4EP4 bugs by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/448
• feat(mla): decode-context-parallel (DCP) support in the MLA decode kernel by @RomaA2000 in https://github.com/lightseekorg/tokenspeed/pull/364
• Update tokenspeed-mla to 0.1.7 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/451
• Update tokenspeed-kernel MLA to 0.1.7 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/452
• Add InstantTensor weight loader (--load-format instanttensor) by @gongwei-130 in https://github.com/lightseekorg/tokenspeed/pull/418
• Update tokenspeed-flashmla to 1.0.0.post20260615 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/458
• docs: move DeepSeek V4 into model recipes (Flash + Pro, MTP) by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/463
• revert: remove instanttensor loader by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/462
• [AMD] Support Kimi K2.5 MXFP4 serving by @qedawkins in https://github.com/lightseekorg/tokenspeed/pull/454
• feat: support trtllm backend features in mha backend by @borontion in https://github.com/lightseekorg/tokenspeed/pull/445
• cleanup: remove unused linear attention helpers by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/467
• refactor(spec-decode): simplify Qwen3.5 NextN attention path for [#217] (2/3) by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/429
• feat(runtime): enhance per-request NaN detect, sanitize, and isolate numerically corrupted requests by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/434
• chore(deps): update smg packages by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/471
• chore(deps): update smg packages by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/475
• feat: adopt trtllm-style spec decode api in mha backend by @borontion in https://github.com/lightseekorg/tokenspeed/pull/465
• chore(deps): update tokenspeed-fa4 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/481
• feat(kernel): register FA4 for SM 103 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/474
• ci: use default mha and moe backend for gpt-oss ci by @borontion in https://github.com/lightseekorg/tokenspeed/pull/479
• perf(Reduce Scatter): Scale reduce-scatter CTA count dynamically by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/457
• ci: use FA4 mm attention for Kimi OCR eval by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/482
• Remove forced NCCL record stream override by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/483
• Use transformers dtype config field by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/484
• fix(deepseek-v4): release superseded interior continuation-state snapshots by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/460
• perf(kernel): Optimize Kimi Vision FA4 QKV + RoPE by @qimcis in https://github.com/lightseekorg/tokenspeed/pull/477
• perf(kernel): Optimize MoE prefill GEMMs for gfx950 by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/464
• Sanitize Qwen3.5 rope parameters by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/485
• fix(log): prevent TP deadlock from log-pipe backpressure under high request-log volume by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/486
• ci: add mi350 1gpu gpt-oss perf bench by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/487
• Update tokenspeed-smg pins to 20260620 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/488
• ci: deduplicate mi350/mi355 gpt-oss bench yaml by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/489
• fix(vlm): concat per-item grids in compute_mrope_positions by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/490
• Update tokenspeed-smg pins to 20260621 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/492
• feat(serve): default ts serve to smg passthrough routing policy by @slin1237 in https://github.com/lightseekorg/tokenspeed/pull/491
• fix(cudagraph): drop bs=0 capture at max_num_seqs=1 by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/493
• Fix gfx950 Gluon MoE preshuffled W2 OOB accesses by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/495
• feat(logprobs): output token logprobs in vLLM and SGLang formats by @HJSang in https://github.com/lightseekorg/tokenspeed/pull/337
• perf(kernel): decode at medium batch sizes by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/470
• Update stale by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/496
• [Model] GLM-5 support: DSA sparse attention, MTP speculative decoding, CUDA graph decode by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/348
• Update smg dependencies by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/497
• chore: make cuda arch configurable by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/498
• feat(video) Generalize multimodal runtime support and add Qwen3.5 video by @yechank-nvidia in https://github.com/lightseekorg/tokenspeed/pull/354
• fix(mla): apply MTP causal mask in BF16/FP16 decode kernel by @zcnrex in https://github.com/lightseekorg/tokenspeed/pull/469
• fix(mla): guard LSE writes when return_lse=False in BF16 decode by @zcnrex in https://github.com/lightseekorg/tokenspeed/pull/468
• fix(cache): widen loc to int64 in MLA KV-buffer Triton kernels to prevent int32 address overflow by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/501
• Update tokenspeed-mla to 0.1.8 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/504
• Update tokenspeed-kernel MLA to 0.1.8 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/505
• Add tokenspeed-kernel-amd release workflow by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/507
• Fix tokenspeed-kernel-amd publishing targets by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/508
• fix(spec): remove V4 MTP special forward modes by @Xiangyi1996 in https://github.com/lightseekorg/tokenspeed/pull/503
• Fix b200 runner env lookup with label overrides by @zhyncs in https://github.com/lightseekorg/tokenspeed/pull/511
• feat(kernel): Tie weight preprocessors to kernels by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/480
• Lower GPQA Diamond eval threshold by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/514
• Fix: GLM5 DSA MTP TP4 IMA Error by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/509
• fix: reuse logits all-gather workspace by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/515
• feat(dflash): Add DFlash support by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/472
• fix(pd): fix Mamba checkpoint handoff by @XucSh in https://github.com/lightseekorg/tokenspeed/pull/516
• Create CODE_OF_CONDUCT by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/517
• Create SECURITY by @torchspec-bot in https://github.com/lightseekorg/tokenspeed/pull/518
• perf(kernel): optimize Qwen vision QKV rotary layout by @qimcis in https://github.com/lightseekorg/tokenspeed/pull/456
• Update README News by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/523
• fix(kernel): out of range offset for kv cache by @borontion in https://github.com/lightseekorg/tokenspeed/pull/522
• Fix stale action cache cleanup by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/524
• Bump tokenspeed-kernel-amd to 0.1.1 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/525
• perf(deepseek-v4): deferred-state MHC forward for cross-layer fusion by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/529
• fix(attention): use int64 paged-KV addressing in MHA Triton kernels by @raikonenfnu in https://github.com/lightseekorg/tokenspeed/pull/530
• feat: distributed argmax for EAGLE greedy sampling by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/519
• Add AMD Kimi MXFP4 CI job by @qedawkins in https://github.com/lightseekorg/tokenspeed/pull/476
• fix(glm5.2): fix indexer_topk_prefil params for batched long prefill IMA by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/533
• fix: decompose PrecisionConfig from tokenspeed-kernel-amd gluon moe by @borontion in https://github.com/lightseekorg/tokenspeed/pull/535
• fix(quant): honor layer-wise quantization exclusion for MoE experts by @jaywme in https://github.com/lightseekorg/tokenspeed/pull/527
• fix(dflash): fix serverl hang issue for dflash and add ci eval by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/531
• Split AMD and NVIDIA PR CI by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/537
• fix(spec): drive EAGLE3 aux capture layers from the draft config by @jaywme in https://github.com/lightseekorg/tokenspeed/pull/526
• fix(sampling): fused_topk_topp PDL race causing IMA by @jaywme in https://github.com/lightseekorg/tokenspeed/pull/536
• ci: temporarily switch to use mi355-1gpu-bench for 1gpu ci by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/538
• ci: switch MI35x back to standard 1GPU CI by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/540
• docs: add blog and talk README link by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/541
• feat(log): add --log-request-stats per-request statistics logging by @LorrinWWW in https://github.com/lightseekorg/tokenspeed/pull/512
• fix(nvfp4): use per-half gate/up global scales for trtllm fused MoE by @jaywme in https://github.com/lightseekorg/tokenspeed/pull/543
• fix: fix cutedsl argmax for NaN inputs and drop now-redundant guards by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/552
• fix(glm-5.2): remove unnecessary DSA KV-cache and flashmla DSA path by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/554
• fix(runtime): harden MTP decode path against NaN, overflow, and state corruption by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/506
• Create CONTRIBUTING by @torchspec-bot in https://github.com/lightseekorg/tokenspeed/pull/559
• Update tokenspeed-smg dependency pins by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/561
• Create Governance by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/564
• fix(scheduler): account for terminal state checkpoints in admission by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/562
• chore(kernel): bump flashinfer to 0.6.13, add <cfloat> include by @jaywme in https://github.com/lightseekorg/tokenspeed/pull/539
• Parallelize scheduler C++ lint by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/566
• fix(scheduler): clamp hybrid TakeFirst page-aligned to prevent OwnedP… by @khushgx in https://github.com/lightseekorg/tokenspeed/pull/565
• Fix gathered MXFP4 activation scales in Gluon MoE by @qedawkins in https://github.com/lightseekorg/tokenspeed/pull/534
• Fix GB200 FlashInfer JIT cache sync by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/567
• Add GLM NVFP4 AIME26 CI by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/569
• fix(ci): resolve GB200 CI FlashInfer JIT cache error by @minedec in https://github.com/lightseekorg/tokenspeed/pull/570
• Initial glm 5.2 support on amd by @borontion in https://github.com/lightseekorg/tokenspeed/pull/528
• feat(glm-5.2): support hierarchical (host/L2) KV cache for DSA by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/572
• fix(sampling): CuteDSL argmax crashes the server on empty-row input by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/545
• fix(attention): emit natural-log LSE from gluon mha prefill by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/571
• Use tokenspeed-spdlog for scheduler builds by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/576
• Fix attention LSE test math import by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/577
• [GLM] perf: Optimize GLM DSA decode top-k by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/573
• fix(attention): treat window_left as exclusive by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/578
• test: glm-5.2 agentic bench by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/532
• Fix DP cache sizing by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/574
• Add Triton sampling backends alongside FlashInfer by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/280
• test: add dp4ep4 case in CI by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/453
• test: fix glm-5.2 agentic bench by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/581
• ci: disable distributed_argmax on qwen3.5 MTP drafter by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/582
• perf(glm-5.2): drop full-topk path and hadamard transform, use flashinfer LayerNorm for DSA decode path by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/586
• Split NVIDIA ARM PR tests by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/588
• Add SPDX source attributions by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/591
• chore(glm-5.2): cleanup bf16 index_k cache for DSA by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/590
• Fix ruff check by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/593
• perf(deepseek-v4): enable MTP overlap scheduling with paged cache by @dongjiyingdjy in https://github.com/lightseekorg/tokenspeed/pull/583
• fix(perf): fix qwen3.5 tp8 agentic ci perf regression by @minedec in https://github.com/lightseekorg/tokenspeed/pull/594
• fix(spec): fix Kimi DP EAGLE3 mixed-step hang by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/596
• feat(attention): add gluon MHA extend kernel for gfx950 by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/587
• feat: breakable CUDA graph for prefill forwards by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/597
• fix(deepseek-v3): Route DeepSeek merge state through unified backend by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/602
• perf(glm-5.2): remove per-token expansion of seq_lens and block_table for DSA decode path by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/599
• Update smg dependency pins by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/606
• fix(kernel): Fix out of bounds accesses in gluon MoE topk implementation by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/605
• Update CODEOWNERS by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/607
• feat(gdn): Initial support for routing chunk prefill through kernel registry by @raikonenfnu in https://github.com/lightseekorg/tokenspeed/pull/595
• fix(deepseek-v4): size CUDA graph sliding-window capture to full window. by @SimonCqk in https://github.com/lightseekorg/tokenspeed/pull/613
• feat(disaggregation): add EPD encode pipeline by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/548
• Add Gluon fused MoE support for dynamic MXFP4 activations by @qedawkins in https://github.com/lightseekorg/tokenspeed/pull/603
• Update smg dependencies by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/615
• Add gfx950 Gluon dense16 GEMM by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/598
• perf(deepseek-v4): sanitize SWA slot mapping once per step by @Xiangyi1996 in https://github.com/lightseekorg/tokenspeed/pull/614
• fix(scheduler): keep hybrid mamba match on device path by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/612
• Update smg dependency pins by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/619
• Add scheduler PyPI publishing by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/623
• Fix scheduler PyPI wheel publishing by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/624
• refactor(spec-decode): simplify deepseekV3/GLM attention path for [#217] (3/3) by @rjzhb in https://github.com/lightseekorg/tokenspeed/pull/544
• ci: drop mi350-1gpu-bench by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/629
• ci: gate tokenspeed-kernel publishing by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/628
• feat(attention): add fp8 support to gfx950 gluon MHA kernels by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/625
• (kernel) : Gluon kernel for BF16 MoE on GFX950 by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/626
• ci: retry failed latest main jobs by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/633
• ci: reenable mi350-1gpu-bench and support optional jobs by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/632
• refactor(attention): consolidate gfx950 gluon MHA layouts into shared helper by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/635
• Update triton to >= 3.8.10.post20260709 by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/636
• ci: retry failed approved PR jobs by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/637
• ci: drop org-membership gate for approved PR reruns by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/638
• Add retry workflow summaries by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/639
• KV-Cache Refactor Part 1: Introduce Unified Block Pool by @wangbo981016 in https://github.com/lightseekorg/tokenspeed/pull/447
• Release tokenspeed-mla 0.1.9 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/641
• Support Heterogeneous block_size by @wangbo981016 in https://github.com/lightseekorg/tokenspeed/pull/647
• feat(attention): trtllm backend consumes flat per-group KV-cache tables by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/642
• Upgrade CuTeDSL to 4.6.0 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/640
• fix(kernel): avoid mixing CUDA toolkit headers by @lucifer1004 in https://github.com/lightseekorg/tokenspeed/pull/644
• Clean up Flat KV-cache tests by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/650
• feat(spec): EAGLE3/MTP speculative decoding on the flat KV-cache path by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/649
• fix(kernel): use Triton FP8 quantization on SM120 by @lucifer1004 in https://github.com/lightseekorg/tokenspeed/pull/646
• build(deps): bump tokenspeed-deepgemm to 2.5.0.post20260629 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/653
• Refresh brand banner to high-res transparent asset by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/656
• Update README by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/657
• feat(multimodal): support ASR and Omni thinker by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/654
• feat: enable the prefill CUDA graph by default by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/611
• [AMD] fix(gluon): support dynamic MXFP4 combine K=256 by @raikonenfnu in https://github.com/lightseekorg/tokenspeed/pull/658
• [GLM] perf: Integrate TRT-LLM cutedsl topk by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/634
• ci: allow excluding unavailable runner labels by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/663
• fix(kernel): Mask gluon MoE tail-k preloads by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/666
• feat(rl): weight-transfer control plane for RL online weight sync by @qywu in https://github.com/lightseekorg/tokenspeed/pull/546
• feat(profile): drive Proton via the /start_profile control plane by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/655
• fix(scheduler): count active pages across groups by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/669
• Add flat KV-cache toggle to scheduler release workflow by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/673
• Bump tokenspeed-scheduler to 0.1.1 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/674
• fix(prefill-graph): illegal memory access on the flat KV-cache path by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/667
• feat(kernel): add a4w4 gluon moe for precomputed top-k by @borontion in https://github.com/lightseekorg/tokenspeed/pull/671
• Route MLA RoPE KV writes through embedding API by @qedawkins in https://github.com/lightseekorg/tokenspeed/pull/643
• refactor(kernel): use runtime router expert counts by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/675
• refactor: drop ad-hoc usage of is hip by @borontion in https://github.com/lightseekorg/tokenspeed/pull/676
• ci: allow configured repositories to run GPU workflows by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/677
• [AMD] Initial support for qwen3.5-397b-a17b-mxfp4 on AMD hardware by @raikonenfnu in https://github.com/lightseekorg/tokenspeed/pull/592
• build(deps): bump tokenspeed-smg packages to post20260714 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/679
• Add quack-kernels to CUDA third-party requirements by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/680
• Add MHA backend scale arguments by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/682
• fix(sampling): fix fused_topk_topp misaligned-address crash without losing float4 throughput by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/683
• perf: flat pagetable export by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/681
• Bump tokenspeed-scheduler to 0.1.2 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/684
• feat(kernel): add gfx950 Gluon mxfp4 MoE decode and prefill kernels by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/670
• Add NVIDIA and AMD runner Dockerfiles by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/685
• ci: fix AMD image namespace by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/686
• feat: support Qwen3.5 DFlash and its optimizations by @minedec in https://github.com/lightseekorg/tokenspeed/pull/510
• Add missing copyright headers by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/687
• Add Inkling recipes by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/690
• Update Inkling link by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/691
• Add Inkling day 0 support news by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/692
• feat(kernel): Add gluon kernel for DSA by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/627
• perf(kernel): tune gfx950 A4W4 MXFP4 MoE for Kimi-K2.5 (decode cap + prefill tile) by @panditsa in https://github.com/lightseekorg/tokenspeed/pull/698
• Add TML Inkling support by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/689
• ci: remove m2 ui by @borontion in https://github.com/lightseekorg/tokenspeed/pull/704
• Update SMG package pins by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/705
• Update tokenspeed-mla to 0.2.0 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/707
• Update tokenspeed-kernel MLA to 0.2.0 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/708
• feat(kernel): Add batch matmul API by @qedawkins in https://github.com/lightseekorg/tokenspeed/pull/668
• ci: add inkling eval ci by @borontion in https://github.com/lightseekorg/tokenspeed/pull/703
• ci: update FlashInfer to 0.6.15 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/713
• fix(inkling): require SMG parsers at startup by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/712
• fix(kvcache): route negative pages to the dummy page in flat_decode_locs by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/709
• fix: cast NVFP4 deferred route weights to bf16 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/716
• feat(kernel): add mxfp MoE Gluon kernel on gfx1250 by @knwng in https://github.com/lightseekorg/tokenspeed/pull/710
• feat(kernel): add gfx950 Gluon bf16 MLA decode and prefill kernels by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/664
• [AMD] Add initial support for pipelined GFX1250 Gluon MHA prefill by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/719
• fix(runtime): restore Qwen3.5 OCRBench accuracy to unblock CI by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/715
• ci(eval): add Qwen3.5 aggregate and EPD OCRBench coverage by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/549
• perf(kernel): optimize gfx950 gluon MHA extend kernel by @Yu-Zhewen in https://github.com/lightseekorg/tokenspeed/pull/717
• fix(profile): barrier TP peers before replying to a Proton /stop_profile by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/722
• Stabilize B200 agentic execution by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/725
• feat(profile): merge Proton and VizTracer traces onto one timeline by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/662
• build: remove duplicate quack-kernels requirement by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/726
• chore: add tokenspeed-kernel-amd copyright headers by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/727
• build: bump tokenspeed kernel packages to 0.1.2 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/728
• build: pin tokenspeed-kernel-amd 0.1.2 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/729
• ci: Reenable mi355x and drop mi350x 1gpu bench by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/732
• refactor(spec): separate multi-layer MTP from Eagle by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/730
• feat(multimodal): add item-level encoder data parallelism by @chenht2022 in https://github.com/lightseekorg/tokenspeed/pull/731
• test(kernel): verify kernel selection cross-platform and drop expected_kernel_name by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/659
• feat(profile): Add multi-rank trace merge support by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/734
• chore: remove legacy spec_decode package and dead ScheduleBatch path by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/739
• refactor(epd): promote epd package and split mooncake transfer module by @FC-Li in https://github.com/lightseekorg/tokenspeed/pull/743
• Update tokenspeed-smg* pins to 1.8.0.post20260720 / 0.4.14.post20260720 / 0.7.0.post20260720 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/744
• [GLM] perf: optimize indexer small kernels by @yweng0828 in https://github.com/lightseekorg/tokenspeed/pull/742
• perf(gdn): use flashinfer decode gdn kernel by @tuanzhangCS in https://github.com/lightseekorg/tokenspeed/pull/700
• feat: add basic MiniMax M3 support by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/733
• feat(m3): support Eagle3.1 draft checkpoints by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/751
• feat(m3): NVFP4 mixed-precision checkpoint support by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/749
• fix(kernel): Widen fused RoPE KV triton kernel cache offsets to i64 by @Max191 in https://github.com/lightseekorg/tokenspeed/pull/748
• fix: clear padded breakable graph handoffs by @FlamingoPg in https://github.com/lightseekorg/tokenspeed/pull/741
• deps: update TokenSpeed Triton dependencies by @antiagainst in https://github.com/lightseekorg/tokenspeed/pull/753
• Update tokenspeed-mla to 0.2.1 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/755
• Update tokenspeed-kernel MLA to 0.2.1 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/754
• Update FlashInfer to 0.6.15.post1 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/757
• ci: add tokenspeed-kernel-amd release lifecycle workflows by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/756
• build: update tokenspeed-kernel-amd to 0.1.3 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/760
• build: sync tokenspeed-kernel with AMD 0.1.3 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/761
• feat(m3): support FP8 KV cache by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/759
• refactor(cache): make BlockRef a shared handle by @wangbo981016 in https://github.com/lightseekorg/tokenspeed/pull/762
• fix(qwen3.5): Fix Qwen3.5 FP8 weight loading by @minedec in https://github.com/lightseekorg/tokenspeed/pull/766
• feat(m3): adopt MiniMax MSA CuTe-DSL sparse prefill attend on SM100 by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/764
• fix(inkling): fuse dMel lookup+sum to bound audio encode memory by @nperrin-fr in https://github.com/lightseekorg/tokenspeed/pull/771
• Bump tokenspeed-scheduler to 0.1.3 by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/772
• ci: remove legacy Docker release files by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/773
• ci: support dated tokenspeed-kernel releases by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/774
• perf(m3): flashinfer cute-dsl backend for MXFP8 (1,32) dense GEMM by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/765
• perf(m3): integrate CuTeDSL MSA block-score kernels by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/775
• ci: disable thinking and allow longer context for qwen3.5 eval/pd cases by @minedec in https://github.com/lightseekorg/tokenspeed/pull/776
• Fix tokenspeed-kernel runtime dependency metadata by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/778
• perf(m3): optimize MiniMax M3 MoE and MSA by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/779
• perf(m3): optimize MiniMax M3 attention module by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/781
• build: declare TokenSpeed runtime dependencies by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/783
• ci: use trusted publishing for PyPI by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/784
• test(m3): enable MiniMax-M3 CI (replacing M2.7) and add agentic benchmark by @syuoni in https://github.com/lightseekorg/tokenspeed/pull/785
• ci: add release Docker publishing by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/788
• ci: add ROCm 7.2 tokenspeed-kernel release workflow by @borontion in https://github.com/lightseekorg/tokenspeed/pull/787
• ci: add TokenSpeed release workflow by @lightseek-bot in https://github.com/lightseekorg/tokenspeed/pull/792

New Contributors

• @zhyncs made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/1
• @antiagainst made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/3
• @knwng made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/4
• @yubofredwang made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/9
• @syuoni made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/17
• @huangzhilin-hzl made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/12
• @andyluo7 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/23
• @XucSh made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/14
• @LorrinWWW made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/35
• @tuanzhangCS made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/15
• @dongjiyingdjy made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/30
• @minedec made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/32
• @roycho96 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/41
• @FlamingoPg made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/10
• @SimonCqk made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/51
• @raayandhar made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/82
• @Aphoh made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/8
• @lite-keke made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/97
• @benenzhu made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/119
• @yweng0828 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/126
• @jasl made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/134
• @elwhyjay made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/155
• @mmangkad made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/163
• @wtsung made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/185
• @curious-whq made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/182
• @yurekami made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/240
• @rjzhb made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/217
• @wangbo981016 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/327
• @Xiangyi1996 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/207
• @Yu-Zhewen made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/313
• @botieking98 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/311
• @HJSang made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/346
• @panditsa made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/363
• @rucnyz made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/397
• @Max191 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/421
• @RomaA2000 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/364
• @gongwei-130 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/418
• @qedawkins made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/454
• @slin1237 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/491
• @yechank-nvidia made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/354
• @zcnrex made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/469
• @torchspec-bot made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/518
• @jaywme made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/527
• @khushgx made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/565
• @nperrin-fr made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/597
• @lucifer1004 made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/644
• @FC-Li made their first contribution in https://github.com/lightseekorg/tokenspeed/pull/743

Full Changelog: https://github.com/lightseekorg/tokenspeed/commits/v0.1.0