fix(mem_cache): alloc_req_slots matches upstream ReqToTokenPool.alloc signature#1

Closed

bhaktatejas922 wants to merge 180 commits into

morph/fix-alloc-reqs-upstream-signature

Commits on Apr 5, 2026

fix: add missing f-string prefixes in warning and assert messages (#22067)

authored
feat: CI auto-bisect workflow for automated regression analysis (#22119)

Kangyan-Zhou
and
claude
authored
[AMD]: Support MLA with nhead<16 and FP8 KV cache for TP=8 (Kimi K2.5… (#21213)

ZiguanWang
and
RoyWang
authored
Remove flaky TestToolChoiceLfm2Moe from test_tool_choice (#22137)
hnyls2002
authored
[CI]Temporary ban auto benchmark tool test (#22138)
Fridge003
authored
Update ci_auto_bisect.py to use correct model (#22142)
Kangyan-Zhou
authored
Add failfast flag to rerun-test workflow (#22141)
hnyls2002
authored
Migrate reasoning_tokens tests to existing server fixtures (#22102)
hnyls2002
authored
[model] support voxtral (speech-to-text) (#21635)

LiYomi
and
mengxiancheng03
authored
[diffusion] fix: fix accuracy for flux series (#22059)

yhyang201
and
mickqian
authored
Consolidate reasoning tests into test/registered/reasoning/ (#22139)
hnyls2002
authored
[Fix] Fix nightly tests (#22140)
Fridge003
authored
Enable IndexCache for DeepSeek V3.2 (#21405)

jinyouzhi
and
Fridge003
authored
[diffusion] Add is_float64_supported to Platform (#22112)
yeahdongcn
authored
[Doc] Update GLM-5 instructions in sglang documentation (#21716)
Fridge003
authored
Isolate spec V1 path in decode post-processing (#22146)
hnyls2002
authored
Add dump_metric to MMMU, lm-eval, and NeMo Skills eval paths (#22147)
hnyls2002
authored
Unify think_end_id to model_config as single source of truth (#22148)
hnyls2002
authored
DEBUG: reproduce flaky test_load_weights_from_remote_instance (#22150)

hnyls2002
and
ShangmingCai
authored
[PD] Fix staging warmup for GQA prefill decode different tp (#22153)
ShangmingCai
authored
[BugFix][RadixTree]: Fix backup invariant violation in Hi-MambaRadixTree (#22062)

hzh0425
and
huangtingwei9988
authored
[SpecV2]: Reopen kl accuracy test for qwen3 + SpecV2 (#22104)
hzh0425
authored
[Hotfix] Fix router gemm on sm103 (#22134)
Fridge003
authored
fix: TRT-LLM MHA CUDA illegal address with EAGLE v2 + DP attention (#21649)

authored
Fix create_grammar_backend test calls with think_end_id (#22158)
ShangmingCai
authored
Update ci_auto_bisect.py to have streak 1 so that all failures will b… (#22161)

Kangyan-Zhou
and
claude
authored
[Misc] Remove unused cu13 docker release workflow (#22167)
Fridge003
authored
Hisparse Minor Fix (#22131)

authored
[Fix] Fix setuptools-scm version resolution for rc tags (#22165)

authored

Commits on Apr 6, 2026

Commits on Apr 7, 2026

Commits on Apr 8, 2026

Commits on Apr 12, 2026

Add support for b12x PCIe oneshot all-reduce backend
lukealonso
committed
Add b12x NSA backend
lukealonso
committed
fix(qwen3_5): extend quant guard for compressed-tensors NVFP4 checkpoints

authored and
voipmonitor build
committed
PR #20534
voipmonitor build
committed
PR #20275
voipmonitor build
committed
Fix AWQ quantization ignoring modules_to_not_convert for FusedMoE layers

voipmonitor
authored and
voipmonitor build
committed
PR #20448
voipmonitor build
committed
PR #20059
voipmonitor build
committed
PR #20460
voipmonitor build
committed
fix(model_runner): warm up cuBLAS/cuBLASLt handles for FP16/BF16

voipmonitor
authored and
voipmonitor build
committed
PR #21601: FP4 KV cache support for SM120 GPUs (NVFP4 E2M1)
voipmonitor build
committed
PR #20479: Triton MLA FP8 KV cache for SM120
voipmonitor build
committed
PR #20433
voipmonitor build
committed
Avoid GPU syncs in mamba track metadata

Chen-0210
authored and
voipmonitor build
committed
PR #21599
voipmonitor build
committed
PR #20114
voipmonitor build
committed
PR #20182
voipmonitor build
committed

Commits on Apr 13, 2026

Add stub classes for PinPrefixReq, LoRA outputs (from unmerged PRs)
voipmonitor build
committed
Revert PR #20182 (mamba memory leak) — scheduler.py incompatible with our utils
voipmonitor build
committed
Fix merge artifact in attention_registry.py (stray -21601 line)
voipmonitor build
committed
Fix PR #21601: take memory_pool.py and model_runner_kv_cache_mixin.py from PR branch
voipmonitor build
committed
Fix modelopt_quant.py: use luke's version (b12x MoE path via apply, not MoeRunner)
voipmonitor build
committed

Commits on Apr 18, 2026

fix: add null check for rid before abort in tokenizer manager

lujangus
authored and
Ubuntu
committed
fix: handle BatchEncoding and prevent double-wrapping in GenerateReqInput

lujangus
authored and
Ubuntu
committed
feat: MiniMax-M2.5 Eagle3 support

lujangus
authored and
Ubuntu
committed
fix: cast aux_hidden_states dtype in Eagle3 draft forward

lujangus
authored and
Ubuntu
committed
fix: complete FP8+Eagle3 dtype compatibility for MoE targets

lujangus
authored and
Ubuntu
committed
fix: EAGLE3 + SWAKVPool compatibility (envs, memory allocator)

lujangus
authored and
Ubuntu
committed
fix(allocator): prevent double-free crash in SWAKVPool + EAGLE3

lujangus
authored and
Ubuntu
committed
fix: CUDA graph crash — use boolean allocated_mask instead of -1 sentinel

lujangus
authored and
Ubuntu
committed
fix: add missing hybrid_kvcache_ratio to ServerArgs

authored and
Ubuntu
committed
fix: add get_rope_config + getattr guard for piecewise CUDA graph

authored and
Ubuntu
committed
fix: alloc_req_slots passes reqs list as need_size

authored and
Ubuntu
committed
fix: handle layer=None in GDN backend forward_extend

authored and
Ubuntu
committed
fix: add is_next_layer_sparse to LayerScatterModes.init_new in glm4_moe.py

authored and
Ubuntu
committed
fix(swa): restore full_to_swa_index_mapping in backup/restore for EAGLE3

authored and
Ubuntu
committed
fix: accept pynccl_use_current_stream in init_model_parallel_group

authored and
Ubuntu
committed
fix: add SWATokenToKVPoolAllocator for SpecForge compatibility

authored and
Ubuntu
committed
fix(qwen2): capture final-layer aux hidden state for EAGLE3 training

authored and
Ubuntu
committed
fix: capture final-layer aux hidden state in all EAGLE3-capable models

authored and
Ubuntu
committed
fix(mem_cache): alloc_req_slots calls ReqToTokenPool.alloc with reqs only
Ubuntu
committed