Karmic beta to canonical (Oct 5): rows of #868, through #982

.lil/changes/vllm-813-withdrawal.json

@@ -0,0 +1,13 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-813-withdrawal",
+  "category": "internal",
+  "summary": "DeepSeek V4 routed-MoE tuning again times candidates without the concurrent shared expert",
+  "models": ["DeepSeek-V4-Flash"],
+  "compatibility": "No user action required.",
+  "details": [
+    "The current routed MoE kernel selects the same serving grid either way; an isolated comparison measured no C1, C8 or prefill difference."
+  ],
+  "pull_requests": [813],
+  "evidence": ["https://github.com/local-inference-lab/vllm/pull/813#issuecomment-5794090484"]
+}

.lil/changes/vllm-829.json

@@ -0,0 +1,18 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-829",
+  "category": "fix",
+  "summary": "MiMo-V2 router logits stay FP32 and the DFlash drafter applies its value scale",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "Backport of vllm-project/vllm#57784 (Zyann7): MiMo's BF16 gate weights produce FP32 router logits, as in the model's reference code, so near-tie expert choices no longer flip.",
+    "The DFlash decoder applies its non-unit value scale exactly once to query-path and cached-context values, which raises draft acceptance."
+  ],
+  "pull_requests": [
+    829
+  ]
+}

.lil/changes/vllm-831.json

@@ -0,0 +1,16 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-831",
+  "category": "fix",
+  "summary": "The SM120 BF16-to-FP32 router matmul compiles under fullgraph compilation",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "The CUDA-graph capture guard, its BF16 allocation and the FP32-output matmul run inside one SM120 custom op, so torch.compile no longer traces a Python bool from the capture query. Other platforms and router paths are unchanged."
+  ],
+  "pull_requests": [
+    831
+  ]
+}

.lil/changes/vllm-832.json

@@ -0,0 +1,18 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-832",
+  "category": "fix",
+  "summary": "--kv-cache-dtype fp8 reaches the MiMo target and MTP attention layers",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "The MiMo target decoder and native MTP layers now receive the cache configuration, so an FP8 KV request allocates FP8 pages instead of BF16 and keeps scale calibration.",
+    "Global layers keep full attention: the model-wide sliding-window fallback is cleared on a copy of the configuration for them."
+  ],
+  "pull_requests": [
+    832
+  ]
+}

.lil/changes/vllm-861.json

@@ -0,0 +1,13 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-861",
+  "category": "fix",
+  "summary": "Keep DCP top-k attention selection in a stable order",
+  "models": ["Qwen3.8-Flash-Next", "GLM-5.3-Flash", "DeepSeek-V4-Flash", "DeepSeek-V4.1-Flash"],
+  "compatibility": "No user action required. Only decode context parallelism (DCP > 1) is affected.",
+  "details": [
+    "DCP candidate compaction preserves the selected order instead of depending on atomic arrival order.",
+    "The merge kernel is shared by the Qwen QSA, GLM and DeepSeek sparse indexers."
+  ],
+  "pull_requests": [861]
+}

.lil/changes/vllm-862.json

@@ -0,0 +1,12 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-862",
+  "category": "performance",
+  "summary": "Optionally bound prefill work per compute-share step under decode load",
+  "models": ["all"],
+  "compatibility": "No user action required.",
+  "details": [
+    "A new step-token limit applies only when prefill and decode contend; the default leaves scheduling unchanged."
+  ],
+  "pull_requests": [862]
+}

.lil/changes/vllm-863.json

@@ -0,0 +1,12 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-863",
+  "category": "fix",
+  "summary": "Start Qwen without a loaded vision stage during B12X warmup",
+  "models": ["Qwen3.8-Flash-Next"],
+  "compatibility": "No user action required.",
+  "details": [
+    "Warmup skips a missing multimodal stage rather than treating its placeholder as a vision module."
+  ],
+  "pull_requests": [863]
+}

.lil/changes/vllm-864.json

@@ -0,0 +1,12 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-864",
+  "category": "feature",
+  "summary": "Allow Qwen QSA checkpoints to use compatible external KV storage",
+  "models": ["Qwen3.8-Flash-Next"],
+  "compatibility": "No user action required.",
+  "details": [
+    "Qwen KV transfer requires a connector that saves and restores an atomic request-boundary checkpoint."
+  ],
+  "pull_requests": [864]
+}

.lil/changes/vllm-865.json

@@ -0,0 +1,14 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-865",
+  "category": "fix",
+  "summary": "Stop Qwen QSA from corrupting the KV cache when CUDA graphs are larger than the decode-row limit",
+  "models": ["Qwen3.8-Flash-Next"],
+  "compatibility": "No user action required. The shipped profile was not affected; deployments that lowered max-num-seqs without lowering the CUDA graph capture size were.",
+  "details": [
+    "Mixed batches replayed in graphs with more rows than max-num-seqs times (1 + draft tokens) could write another request's compressed KV page, causing wrong output without an error, or an illegal memory access.",
+    "QSA now sizes captured graphs for the full model length."
+  ],
+  "pull_requests": [865],
+  "evidence": ["https://github.com/local-inference-lab/vllm/pull/865"]
+}

.lil/changes/vllm-866.json

@@ -0,0 +1,14 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-866",
+  "category": "feature",
+  "summary": "Log where a request or the engine-core loop is held",
+  "models": ["all"],
+  "compatibility": "No user action required.",
+  "details": [
+    "A warning names the stage when a request spends more than VLLM_REQUEST_STALL_WARNING_S (default 60) outside engine work, or ends without output after that long; VLLM_LOG_REQUEST_TIMELINE=1 logs every request's stages.",
+    "The engine core warns about ADD requests that waited in its input queue and logs the loop stack when one step runs longer than the threshold.",
+    "SIGUSR1 to the API server logs every asyncio task's await chain and every thread's stack, without prompt or output text."
+  ],
+  "pull_requests": [866]
+}

.lil/changes/vllm-869.json

@@ -0,0 +1,19 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-869",
+  "category": "fix",
+  "summary": "Track allocated KV blocks and keep CoW group store ranges aligned in the offloading connector",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "Two OffloadingConnector fixes from #867: newly allocated blocks are tracked before an early return (port of vllm-project/vllm#46960), and skipped align-mode CoW groups keep their store-range slots, which stops an EngineCore assert on the first store of a long prompt."
+  ],
+  "pull_requests": [
+    869
+  ],
+  "authors": [
+    "@llitz"
+  ]
+}

.lil/changes/vllm-870.json

@@ -0,0 +1,21 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-870",
+  "category": "fix",
+  "summary": "Serve external prefix hits for align-mode Mamba hybrids through the offloading connector",
+  "models": [
+    "Qwen3.8-Flash-Next"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "Boundary states are captured in dedicated blocks and QSA pages transfer their padded selector tail, so OffloadingConnector prefix hits on Qwen3.8 are correct.",
+    "A restored boundary checkpoint keeps its QSA selector pools instead of treating the slot as fresh; GPU prefix-hit and LMCache restore answers were unchanged in our Qwen TP2 MTP3 checks.",
+    "Qwen still rejects OffloadingConnector until a connector declares atomic support."
+  ],
+  "pull_requests": [
+    870
+  ],
+  "authors": [
+    "@llitz"
+  ]
+}

.lil/changes/vllm-871.json

@@ -0,0 +1,21 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-871",
+  "category": "feature",
+  "summary": "Serve MiMo V2 decode and speculative verify through b12x paged_decode",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "The B12X backend serves DiffKV layers and non-causal FP8-KV drafter layers through b12x paged_decode, including ragged adaptive-verification batches, and appends their KV through the same plan.",
+    "Other layer geometries keep their existing B12X paths (VLLM_B12X_PAGED_DECODE=auto)."
+  ],
+  "pull_requests": [
+    871
+  ],
+  "requires": [
+    "b12x-420"
+  ]
+}

.lil/changes/vllm-872.json

@@ -0,0 +1,18 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-872",
+  "category": "performance",
+  "summary": "Fuse the b12x PCIe all-reduce with the residual RMSNorm for MiMo V2 and DFlash",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "Fused all-reduce + RMSNorm plans are keyed by the norm weight's storage, so the compiled graph's parameter aliases find them.",
+    "MiMo V2 and the DFlash drafter declare their fused sites; other models are unchanged."
+  ],
+  "pull_requests": [
+    872
+  ]
+}

.lil/changes/vllm-873.json

@@ -0,0 +1,16 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-873",
+  "category": "performance",
+  "summary": "Optionally serve decode-sized BF16 linears through b12x bf16_gemv",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "VLLM_B12X_BF16_GEMV=1 routes BF16 linears with up to eight rows to autotuned b12x GEMV plans under --linear-backend b12x. Default off."
+  ],
+  "pull_requests": [
+    873
+  ]
+}

.lil/changes/vllm-874.json

@@ -0,0 +1,17 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-874",
+  "category": "fix",
+  "summary": "Load MiMo V2 fused FP8 QKV exactly below TP4",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "At TP1/TP2 the fused QKV kept the checkpoint's FP8 block scales instead of being re-quantized (which changed 40% of the global layers' K and 65% of their V weights at TP2). TP4 is unchanged."
+  ],
+  "pull_requests": [
+    874
+  ]
+}

.lil/changes/vllm-875.json

@@ -0,0 +1,17 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-875",
+  "category": "performance",
+  "summary": "Optional L2 weight prefetch for MiMo V2 decode",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "VLLM_MIMO_L2_PREFETCH=1 prefetches upcoming decode weights into L2 on a side stream inside FULL CUDA-graph decode. Cache hints only; numerics unchanged. Default off."
+  ],
+  "pull_requests": [
+    875
+  ]
+}

.lil/changes/vllm-876.json

@@ -0,0 +1,18 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-876",
+  "category": "fix",
+  "summary": "MiMo-V2 audio input works without torchaudio, with identical features",
+  "models": [
+    "XiaomiMiMo/MiMo-V2.6-Flash-RL",
+    "XiaomiMiMo/MiMo-V2.6-Pro-RL"
+  ],
+  "compatibility": "No user action required. Deployments with torchaudio installed are unchanged.",
+  "details": [
+    "Images without a matching torchaudio wheel rejected every MiMo audio request: the media parser's default resampler and the MiMo processor's mel spectrogram both required torchaudio.",
+    "Plain-PyTorch ports of torchaudio's default Resample and MelSpectrogram now serve when torchaudio is missing; they are bitwise identical to torchaudio (tests compare with torch.equal when torchaudio is present and pin reference values when it is not)."
+  ],
+  "pull_requests": [
+    876
+  ]
+}

.lil/changes/vllm-877.json

@@ -0,0 +1,17 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-877",
+  "category": "performance",
+  "summary": "Shard the DFlash context projection and optionally draft per vocab shard at TP > 1",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "The DFlash context projection is column-parallel at TP > 1 (bit-identical output, 1/tp of the weight per rank).",
+    "VLLM_DFLASH_VOCAB_PARALLEL_DRAFT=1 samples probabilistic drafts per vocab shard with identical draws and verification, without the draft-logits all-gather. Default off."
+  ],
+  "pull_requests": [
+    877
+  ]
+}

.lil/changes/vllm-878.json

@@ -0,0 +1,16 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-878",
+  "category": "fix",
+  "summary": "Support block-FP8 DFlash drafters in the fused context projection",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "Online fp8_per_block drafters failed in the fused context K/V projection (BF16 x FP8 matmul); their K/V rows are now dequantized with their block scales for that projection."
+  ],
+  "pull_requests": [
+    878
+  ]
+}

.lil/changes/vllm-879.json

@@ -0,0 +1,16 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-879",
+  "category": "fix",
+  "summary": "Keep adaptive verification's acceptance estimator from collapsing",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "One confidently wrong draft could drive the estimator's predicted acceptance to ~1e-8, after which adaptive verification admitted no drafts for the rest of the run (DFlash or DSpark drafting stopping mid-generation). Refits are now bounded and an idle-step probe keeps evidence flowing."
+  ],
+  "pull_requests": [
+    879
+  ]
+}

.lil/changes/vllm-880.json

@@ -0,0 +1,16 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "vllm-880",
+  "category": "performance",
+  "summary": "Optionally let prefill-only scheduling steps use the full batched-token budget",
+  "models": [
+    "all"
+  ],
+  "compatibility": "No user action required.",
+  "details": [
+    "VLLM_SCHEDULER_UNCAP_PREFILL_ONLY_STEPS=1 lets steps with no runnable decode prefill up to max_num_batched_tokens. Default off."
+  ],
+  "pull_requests": [
+    880
+  ]
+}