[beta] Integrate CSF preparation, native expert precision and decoder optimizations

.lil/changes/b12x-450-csf-canonical.json

@@ -0,0 +1,21 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-450-csf-canonical",
+  "category": "feature",
+  "summary": "Load canonical MXFP4-CSF and NVFP4-CSF checkpoints, including Qwen3.8-Flash-Next",
+  "models": [
+    "Kimi-K3",
+    "DeepSeek-V4.1-Flash",
+    "GLM-5.3-Flash",
+    "Qwen3.8-Flash-Next"
+  ],
+  "compatibility": "Requires CSF schema, tensor suffixes, and loader names. Predecessor X4T/LSC schemas require migration. Original FP4 values and calibration are preserved.",
+  "details": [
+    "TP-independent checkpoint slices preserve native expert arithmetic.",
+    "NVFP4-CSF stores resident offsets in native scale order and initializes micro-kernel barriers during scale expansion."
+  ],
+  "pull_requests": [
+    450
+  ],
+  "requires": []
+}

.lil/changes/b12x-450-csf-mxfp8.json

@@ -0,0 +1,15 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-450-csf-mxfp8",
+  "category": "fix",
+  "summary": "Prepare MXFP4-CSF scales for native MXFP8 expert activations",
+  "models": ["DeepSeek-V4.1-Flash", "GLM-5.3-Flash", "Qwen3.8-Flash-Next"],
+  "compatibility": "Canonical checkpoint bytes are unchanged. Pair with the vLLM DS4.1 activation-policy update. Kimi retains BF16 expert activations.",
+  "details": [
+    "MXFP4-CSF supports the native W4A8 prepared scale layout, including compact N64 tails, native E8M0 clamping and caller-owned shared scratch.",
+    "Compressed-scale expansion skips absent experts using a bounded presence mask instead of repeatedly scanning route prefixes or assuming every expert is active.",
+    "The DS4.1 TP4 batch execution-rate regression is reduced to approximately 1.4 percent or less at concurrency 8 through 32 in the measured configuration. NVFP4 serving retains a material regression; its complete repair is not claimed."
+  ],
+  "pull_requests": [450],
+  "requires": ["b12x-450-csf-preparation"]
+}

.lil/changes/b12x-450-csf-preparation.json

@@ -0,0 +1,22 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-450-csf-preparation",
+  "category": "compatibility",
+  "summary": "Prepare compressed scale planes through the fused-MoE weight API",
+  "models": [
+    "Kimi-K3",
+    "DeepSeek-V4.1-Flash",
+    "GLM-5.3-Flash",
+    "Qwen3.8-Flash-Next"
+  ],
+  "compatibility": "Install the paired vLLM/B12X CSF preparation changes. Canonical CSF checkpoint files and launch commands remain compatible.",
+  "details": [
+    "CsfScalePlanes carries rank-local CPU fixed and exception tensors; prepare_weights creates resident scale batches and retains the ordinary weight and scale preparation paths.",
+    "CSF checkpoint readers and tensor-store interfaces are removed from B12X; vLLM owns file access, model mapping and TP slicing.",
+    "Scale decoders, expert kernels and shared scratch ownership are unchanged."
+  ],
+  "pull_requests": [
+    450
+  ],
+  "requires": []
+}

.lil/changes/b12x-450-csf-tensor-sources.json

@@ -0,0 +1,15 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-450-csf-tensor-sources",
+  "category": "compatibility",
+  "summary": "Load compressed FP4 expert tensors without model-specific checkpoint readers in B12X",
+  "models": ["Kimi-K3", "DeepSeek-V4.1-Flash", "GLM-5.3-Flash", "Qwen3.8-Flash-Next"],
+  "compatibility": "Requires the companion vLLM tensor-source loader. The B12X root/path reader API is replaced; canonical CSF checkpoint bytes and serving flags are unchanged.",
+  "details": [
+    "CsfMatrix sources carry lazy packed weight views, compressed scales and optional FP32 NVFP4 calibration.",
+    "B12X performs generic TP slicing and GPU preparation; vLLM owns manifests, model tensor names and file handles.",
+    "Compute kernels, scale layouts and caller-owned scratch reuse are unchanged."
+  ],
+  "pull_requests": [450],
+  "requires": []
+}

.lil/changes/b12x-450-moe-routing.json

@@ -0,0 +1,15 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-450-moe-routing",
+  "category": "fix",
+  "summary": "Preserve route broadcasts and shared-input token addressing in dynamic MoE",
+  "models": [],
+  "compatibility": "Applies to native and compressed weights. Deterministic W4A8 M32 execution now reads the correct token activation for each routed pair.",
+  "details": [
+    "A CTA barrier prevents reuse of the shared work-claim slot before every warp has retained its value.",
+    "Shared W4A8 payload and scale gathers convert deterministic routed-pair indices back to token indices before computing Int64 global offsets.",
+    "Preparation retains the separate deterministic top-k reduction as well as every selected scale decoder before frozen graph replay."
+  ],
+  "pull_requests": [450],
+  "requires": []
+}

.lil/changes/b12x-450-nvfp4-csf-cooperative.json

@@ -0,0 +1,15 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-450-nvfp4-csf-cooperative",
+  "category": "performance",
+  "summary": "Reconstruct NVFP4-CSF scale operands inside dynamic expert kernels",
+  "models": ["GLM-5.3-Flash", "Qwen3.8-Flash-Next"],
+  "compatibility": "Canonical checkpoint bytes, FP32 calibration and A4 arithmetic are unchanged. Preparation retunes its versioned decoder choice; no launch flag changes are required.",
+  "details": [
+    "Typed preparation selects complete-plane expansion or cooperative reconstruction in shared operand slots. N64 intermediate tails are supported.",
+    "Short-route expansion uses equal-size output spans and the prepared exception-word index. Scratch remains caller-owned and graph replay does not allocate.",
+    "The serving report records measured concurrency, hardware, source identities and remaining overhead; component decoder timings alone do not establish model throughput."
+  ],
+  "pull_requests": [450],
+  "requires": ["b12x-450-csf-mxfp8"]
+}

.lil/changes/b12x-456-block-codec-shared-memory.json

@@ -0,0 +1,14 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-456-block-codec-shared-memory",
+  "category": "fix",
+  "summary": "Reserve Trellis lookup-table shared memory only for Trellis weights",
+  "models": ["all"],
+  "compatibility": "No user action required.",
+  "details": [
+    "Block-codec and ordinary FP4 expert kernels do not allocate an unused Trellis codebook table.",
+    "IQ2_XS and Q8_0 tuning candidates retain the shared-memory capacity checked by their preparation contract; Trellis table selection is unchanged."
+  ],
+  "pull_requests": [456],
+  "requires": []
+}

.lil/changes/b12x-456-csf-tuning-contract.json

@@ -0,0 +1,14 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-456-csf-tuning-contract",
+  "category": "compatibility",
+  "summary": "Retain native MXFP8 and Trellis tuning with compressed FP4 scale operands",
+  "models": ["all"],
+  "compatibility": "No user action required.",
+  "details": [
+    "MoE tuning preserves native MXFP8 backend restrictions and Trellis decode-table validation while accepting NVFP4-CSF operand reconstruction.",
+    "Separate query and configuration schema versions invalidate cached choices from incompatible recipe contracts."
+  ],
+  "pull_requests": [456],
+  "requires": ["b12x-450-nvfp4-csf-cooperative"]
+}

.lil/changes/b12x-csf-indexed-vectors.json

@@ -0,0 +1,15 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-csf-indexed-vectors",
+  "category": "performance",
+  "summary": "Share NVFP4-CSF address and exception lookup across four adjacent scale words",
+  "models": ["Qwen3.8-Flash-Next", "GLM-5.3-Flash"],
+  "compatibility": "Checkpoint schemas, prepared scale storage, activation arithmetic and decoder launch geometry are unchanged.",
+  "details": [
+    "Load eight packed code bytes and four row bases together, then store sixteen reconstructed scale bytes.",
+    "Use one bitmap and prefix lookup per four-word group while preserving individual exception words.",
+    "Keep expert-scaled addresses in 64-bit arithmetic, including output regions beyond 4 GiB."
+  ],
+  "pull_requests": [459],
+  "requires": ["b12x-csf-operand-reconstruction"]
+}

.lil/changes/b12x-csf-operand-reconstruction.json

@@ -0,0 +1,22 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-csf-operand-reconstruction",
+  "category": "performance",
+  "summary": "Reduce synchronization and redundant reads in NVFP4-CSF scale reconstruction",
+  "models": [
+    "Qwen3.8-Flash-Next",
+    "GLM-5.3-Flash"
+  ],
+  "compatibility": "Checkpoint schemas, activation precision, preparation API and scale storage sizes are unchanged.",
+  "details": [
+    "Reconstruct ready gate/up operands with one pair of consumer barriers.",
+    "Reconstruct only the consumed 64-row halves when a gate crosses 128-row scale atoms.",
+    "Select exact zero padding for partial K atoms inside the shared scale reader."
+  ],
+  "pull_requests": [
+    459
+  ],
+  "requires": [
+    "b12x-450-nvfp4-csf-cooperative"
+  ]
+}

.lil/changes/b12x-dense-mla-fp32-partials.json

@@ -0,0 +1,18 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-dense-mla-fp32-partials",
+  "category": "feature",
+  "summary": "Allow dense MLA split partials to remain FP32 until the output merge",
+  "models": [
+    "Kimi-K3"
+  ],
+  "compatibility": "Set Caps.partial_dtype=torch.float32; torch.bfloat16 remains the default.",
+  "details": [
+    "FP32 doubles partial-output scratch without changing KV dtype, output dtype or split-selection policy."
+  ],
+  "pull_requests": [],
+  "requires": [],
+  "authors": [
+    "voipmonitor"
+  ]
+}

.lil/changes/b12x-kimi-x4t-prepared-scales.json

@@ -0,0 +1,16 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-kimi-x4t-prepared-scales",
+  "category": "feature",
+  "summary": "Prepare lossless Kimi-K3 X4T experts and retain compressed-scale programs for frozen serving.",
+  "models": [
+    "Kimi-K3"
+  ],
+  "compatibility": "Adds Kimi TP1/2/4/8/12/16 geometry to the DS4.1 X4T reader. TP16 serving is measured; other topologies require separate qualification.",
+  "pull_requests": [
+    446
+  ],
+  "authors": [
+    "voipmonitor"
+  ]
+}

.lil/changes/b12x-master-sync-20261001.json

@@ -0,0 +1,19 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-master-sync-20261001",
+  "category": "performance",
+  "summary": "Carry current B12X master into beta: shared Trellis/EXL3 expert preparation with full-codeword K2 decoding",
+  "models": [
+    "Kimi-K3"
+  ],
+  "compatibility": "No user action required. The beta keeps version 1.3.0, which vLLM's b12x extra pins; master's 1.5.0 release bump is not carried.",
+  "details": [
+    "Uniform EXL3 and native Trellis experts share one preparation path, and prepared K2 Trellis decoding looks up full codewords (b12x #436). It removes the 82.56 tok/s regression of the uniform two-bit Kimi-K3 EXL3 checkpoint with unchanged acceptance.",
+    "The same sync brings FP32 dense-MLA split partials as an explicit option (b12x #441) and retained Trellis route-pack programs (b12x #442), described in their own notes."
+  ],
+  "authors": [
+    "lukealonso"
+  ],
+  "pull_requests": [],
+  "requires": []
+}

.lil/changes/b12x-online-csf-scale-preparation.json

@@ -0,0 +1,20 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-online-csf-scale-preparation",
+  "category": "feature",
+  "summary": "Compress native FP4 expert block scales losslessly during weight preparation",
+  "models": [
+    "Kimi-K3",
+    "DeepSeek-V4.1-Flash",
+    "GLM-5.3-Flash",
+    "Qwen3.8-Flash-Next"
+  ],
+  "compatibility": "Opt-in scale_compression=csf requires caller-owned serialized scale scratch. Default preparation and checkpoint codecs are unchanged.",
+  "details": [
+    "GPU byte packing preserves native scale bytes and activation arithmetic; no per-request encoding."
+  ],
+  "pull_requests": [
+    458
+  ],
+  "requires": []
+}

.lil/changes/b12x-preparation-launch-ownership.json

@@ -0,0 +1,11 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-preparation-launch-ownership",
+  "category": "fix",
+  "summary": "Retain compiled collective launchers and coordinate cross-rank priming",
+  "models": ["all"],
+  "compatibility": "No user action required.",
+  "details": ["Preserves scoped program-cache reclamation and shared RoCE compile results, and rejects unsupported RoCE dtypes before launcher construction."],
+  "pull_requests": [384],
+  "requires": []
+}

.lil/changes/b12x-trellis-route-pack-ownership.json

@@ -0,0 +1,15 @@
+{
+  "schema": "local-inference-release-change/v1",
+  "id": "b12x-trellis-route-pack-ownership",
+  "category": "fix",
+  "summary": "Retain full-rotation Trellis route-pack programs in materialized MoE plans",
+  "models": ["Kimi-K3"],
+  "compatibility": "No user action required.",
+  "details": [
+    "Transfers compiled route-pack carriers from the preparation result to caller-owned scratch plans.",
+    "Tests carrier identity, numerical parity, poisoned scratch, and CUDA Graph replay at capacities 1, 8, and 129 with uneven expert extents."
+  ],
+  "pull_requests": [],
+  "authors": ["voipmonitor"],
+  "requires": []
+}

b12x/_lib/compile_plan.py

@@ -16,6 +16,7 @@
 from contextvars import ContextVar
 from dataclasses import dataclass, field
 from typing import Any
+from types import FunctionType
 
 
 @dataclass(frozen=True)
@@ -203,10 +204,12 @@ def launch_triton(kernel, grid, *args, **kwargs):
 
 
 def planning() -> bool:
+    """Report whether the caller is collecting compiler program identities."""
     return _PLANNING.get() is not None
 
 
 def record_program(program: ProgramKey, owner: Any = None) -> None:
+    """Record a compiled program identity for active planning observers."""
     captured = _PLANNING.get()
     if captured is not None:
         captured.add(program)
@@ -219,6 +222,7 @@ def record_program(program: ProgramKey, owner: Any = None) -> None:
 
 
 def program_keys(value: Any) -> tuple[ProgramKey, ...]:
+    """Return compiler program identities retained by ``value``."""
     if value is None or isinstance(value, (str, int, float, bool)) or type(value).__module__ == "torch":
         return ()
     if isinstance(value, Mapping):
@@ -228,6 +232,11 @@ def program_keys(value: Any) -> tuple[ProgramKey, ...]:
     keys = getattr(value, "__b12x_programs__", None)
     if keys is not None:
         return tuple(keys)
+    if isinstance(value, FunctionType):
+        # A launch closure produced outside a compile factory carries no
+        # program identity of its own; retain it rather than failing an
+        # unrelated sweep.
+        return ()
     from triton.compiler.compiler import CompiledKernel
     if isinstance(value, CompiledKernel):
         return (ProgramKey("triton", value.hash, value.name),)