Package GLM MTP3 compute fixes and stream-ordered mesh transport

.gitattributes

@@ -1,6 +1,8 @@
 *.sh text eol=lf
 *.patch text eol=lf
 runtime/** text eol=lf
+runtime/glm53-spark-mtp3-mesh/compute/*.tar.gz -text
+runtime/glm53-spark-mtp3-mesh/compute/*.patch whitespace=-trailing-space
 runtime/exl3/patches/*.patch whitespace=-trailing-space
 
 # Runtime receipts hash these source files byte-for-byte before copying them,

README.md

@@ -51,7 +51,7 @@ by that profile.
 | Image family / profile | Purpose | Start here |
 |---|---|---|
 | [GLM-5.3 DFlash2/SIRCL](runtime/glm53-flash-jj-r8-gb10/README.md) | Linux/ARM64 vLLM image with B12X kernels, DFlash2, SIRCL, and optional SparkCache. | [DFlash2 quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
-| [GLM-5.3 native-MTP3 mesh](runtime/glm53-spark-mtp3-mesh/public-image.json) | Linux/ARM64 image for NVFP4-Spark, native MTP3, hybrid mesh transport, and SparkCache. | [Mesh quickstart](docs/GLM53_SPARK_MTP3_MESH_QUICKSTART.md) |
+| [GLM-5.3 native-MTP3 mesh](runtime/glm53-spark-mtp3-mesh/public-image.json) | Linux/ARM64 image for NVFP4-Spark, CUDA 13.3, native MTP3 with an NVFP4/BF16 proposal head, hybrid mesh transport, and SparkCache. | [Mesh quickstart](docs/GLM53_SPARK_MTP3_MESH_QUICKSTART.md) |
 | [`sparkring-glm53-runtime`](https://github.com/users/FujitsuPolycom/packages/container/package/sparkring-glm53-runtime) | Pinned GLM-5.3 bases used to build serving images. | Use the digest named by the source-build guide. |
 | [`gb10-vllm-serving`](https://github.com/users/FujitsuPolycom/packages/container/package/gb10-vllm-serving) | Profile-specific GB10 images, including DeepSeek. | Use the image named by the selected model quickstart. |
 
@@ -69,18 +69,24 @@ together. In the tables, 1M means 1,048,576 tokens.
 
 ### GLM-5.3 Flash
 
-These three **DFlash2/SIRCL profiles** use the GLM-5.3 Flash NVFP4 target,
-an external BF16 DFlash2 predictor at depth seven, FP8 KV, B12X kernels, and
-the same ARM64 image. Target verification captures use rows 8 through 128 in
-eight-row increments, covering full request batches from C1 through C16.
+Choose native MTP3 with hardware-forwarded mesh for the NVFP4-Spark
+checkpoint, or an external DFlash2 predictor with SIRCL for the NVFP4
+checkpoint. Each quickstart supplies its own pinned image and setup steps.
 
 | Profile | Deployment | Context | Seqs | Batch | KV / cache | Approx. recorded KV capacity | Start here |
 |---|---|---:|---:|---:|---|---:|---|
-| DCP1 | 4 Sparks · TP4/DCP1 | 1M | 16 | 8,192 | FP8 · 24 GiB/rank; SparkCache enabled | ~1.30M tokens | [Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
-| DCP2 | 4 Sparks · TP4/DCP2 | 1M | 16 | 8,192 | FP8 · 24 GiB/rank; SparkCache enabled | ~2.90M tokens | [Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
-| **DCP4 preferred** | **4 Sparks · TP4/DCP4** | **1M** | **16** | **8,192** | **FP8 · 24 GiB/rank; SparkCache enabled** | **~4.32M tokens** | **[Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md)** |
+| **NVFP4-Spark · native MTP3 · mesh** (research-only) | **4 Sparks · TP4/DCP4** | **1M** | **16** | **8,192** | **FP8 · 24 GiB/rank; SparkCache enabled** | Reported at startup | **[Mesh quickstart](docs/GLM53_SPARK_MTP3_MESH_QUICKSTART.md)** |
+| DFlash2/SIRCL · DCP1 | 4 Sparks · TP4/DCP1 | 1M | 16 | 8,192 | FP8 · 24 GiB/rank; SparkCache enabled | ~1.30M tokens | [Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
+| DFlash2/SIRCL · DCP2 | 4 Sparks · TP4/DCP2 | 1M | 16 | 8,192 | FP8 · 24 GiB/rank; SparkCache enabled | ~2.90M tokens | [Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
+| DFlash2/SIRCL · DCP4 (preferred DFlash2 profile) | 4 Sparks · TP4/DCP4 | 1M | 16 | 8,192 | FP8 · 24 GiB/rank; SparkCache enabled | ~4.32M tokens | [Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
 
-All three default to 24 GiB of KV memory per rank. The reference capacities
+#### DFlash2/SIRCL configuration
+
+The three DFlash2 profiles use an external BF16 predictor at depth seven,
+FP8 KV, B12X kernels, and the same ARM64 image. Target verification captures
+use rows 8 through 128 in eight-row increments, covering C1 through C16.
+
+All profiles above default to 24 GiB of KV memory per rank. The DFlash2 reference capacities
 were measured at 26/30/24 GiB for DCP1/2/4 respectively; vLLM reports the
 actual model-wide capacity at startup. That capacity is shared across requests.
 **DCP4 is the preferred DFlash2 profile** and the documented asynchronous
@@ -109,16 +115,25 @@ for its terms. The native-MTP3 profile below does not require these weights.
 
 ### GLM-5.3 Flash Spark with native MTP3 and mesh transport
 
-| Profile | Deployment | Context | Seqs | Batch | KV / cache | Start here |
-|---|---|---:|---:|---:|---|---|
-| NVFP4-Spark + native MTP3 + mesh · research-only | 4 Sparks · TP4/DCP4 | 1M | 16 | 8,192 | FP8 · 24 GiB/rank; SparkCache enabled | [Quickstart](docs/GLM53_SPARK_MTP3_MESH_QUICKSTART.md) |
-
 This profile uses the NVFP4-Spark checkpoint's built-in three-token predictor;
-no external draft checkpoint is required. It combines graph-native SIRCL for
+no external draft checkpoint or DFlash model is required. A separate runtime-
+NVFP4 proposal head uses BF16 activations while the target/verifier head retains
+its BF16 checkpoint representation. The profile combines graph-native SIRCL for
 selected decode shapes, dual-rail SIRCL for large prefill collectives, and
 RoCEnante for selected small all-reduces. Target verification captures use
 four-row increments through 64 rows.
 
+The image composition uses CUDA 13.3, the native-MTP3 metadata port derived
+from Local Inference Lab vLLM revision `3512b066`, and the complete B12X tree at
+`ef308bac` with source-checked top-k selector files and vLLM integration based
+on `a8c796f3`. Deferred weights own their storage, and draft/rejection sampling
+uses independent randomness. The complete B12X update
+also includes MoE and dense-precision work, so comparisons across different
+compute configurations cannot attribute a gain to dense kernels alone. The
+[head-specific comparison](performance/records/glm53-flash/spark-mtp3-nvfp4-proposal-head-20260905.md)
+uses a control with the same CUDA version, B12X kernels, metadata reuse, and
+dense-kernel integration; only the proposal-head configuration differs.
+
 The [profile package](runtime/glm53-spark-mtp3-mesh/README.md) provides the
 public image, transport files, and temperature-one warmup. The
 [mesh service](runtime/glm53-spark-mtp3-mesh/MANAGED_MESH.md) sets up and
@@ -155,7 +170,7 @@ prompt. Each linked record gives its sampling, cache settings, and repeat counts
 
 | Profile | Decode context | Prefill | C1 decode | C8 decode | Highest C at this context | Coding peak |
 |---|---:|---:|---:|---:|---:|---:|
-| [GLM-5.3 NVFP4-Spark · native MTP3 + mesh · 4 Sparks](performance/records/glm53-flash/spark-mtp3-mesh-20260905.md) | 8K | 2,703 (8K scout) | 48.2 | 168.8 | C16: 231.3 | — |
+| [GLM-5.3 NVFP4-Spark · native MTP3 + NVFP4/BF16 proposal head + mesh · 4 Sparks](performance/records/glm53-flash/spark-mtp3-nvfp4-proposal-head-20260905.md) | 8K | 2,670 (8K scout mean) | 51.6 | 168.8 | C8: 168.8 | — |
 | [GLM-5.3 NVFP4-Spark · DFlash2 exact request-batch graphs · 4 Sparks](performance/records/glm53-flash/dflash2-exact-concurrency-graphs-20260904.md) | 16K | 2,717 (16K scout) | 43.05 | 134.3 | C16: 187.0 | — |
 | [GLM-5.3 NVFP4 · DFlash2/B12X-KDA DCP4 · 4 Sparks](performance/records/glm53-flash/b12x-kda-dcp4-20260903.md) | 16K | 2,649 (16K scout) | 37.97 | — | C4: 90.36 | — |
 | [GLM-5.2 EXL3 3.5-bpw · 4 Sparks](performance/records/glm-3.5bpw/normalized-base-20260822.md) | 16K | 671 (16K) | 20.15 | 64.13 | C8: 64.13 | 25.39 |
@@ -164,11 +179,15 @@ prompt. Each linked record gives its sampling, cache settings, and repeat counts
 | [Qwen3.8-27B EXL3 K5/K6 · 2 Sparks](performance/records/qwen38-27b/normalized-tp2-1m-probmtp-temp1-20260823.md) | 16K | 1,367 (16K) | 29.50 | 142.20 | C16: 184.39 | 39.95 |
 | [Qwen3.8-27B EXL3 K5/K6 · 4 Sparks](performance/records/qwen38-27b/normalized-tp4-1m-probmtp-temp1-20260823.md) | 16K | 1,964 (16K) | 35.07 | 191.02 | C8: 191.02 | 48.46 |
 
-The native-MTP3 mesh row uses one observation per cell with caching enabled.
-Its [consolidated report](performance/records/glm53-flash/spark-mtp3-validation-summary-20260905.md)
-also includes three-pass cold-prefix prefill results, the 32K/64K decode
-matrices, Estonia **30/30** at C8, and **4/4** needle-hunt checks through
-507,367 prompt tokens.
+The native-MTP3 proposal-head row uses three observations per decode cell and
+three prefill scouts per listed context. Relative to two shared-BF16-head
+controls with the same compute composition, C1 changed by +8.22% raw output
+throughput and +4.90% normalized sequence steps/s; C2/C4/C8 were mixed and
+prefill was flat within 0.36%. The image-specific
+[consolidated report](performance/records/glm53-flash/spark-mtp3-validation-summary-20260905.md)
+retains the 32K/64K matrix, Estonia **30/30** at C8, and **4/4** needle-hunt
+checks through 507,367 prompt tokens, with the image identity recorded for
+each test.
 
 See [benchmark results and throughput tables](docs/RESULTS.md) for full
 matrices, sample counts, exact settings, and limitations.

THIRD_PARTY_NOTICES.md

@@ -55,7 +55,17 @@ two compatibility patches. The public GLM-5.3 builder does not consume them;
 its `NCCL_SWITCHLESS_RING_ONLY` parameter and diagnostics are an independent
 SparkRing implementation.
 
-## 3. vLLM (referenced and patched)
+## 3. vLLM (referenced, patched, and selected source included)
+
+`runtime/glm53-spark-mtp3-mesh/compute/vllm-compute-files.tar.gz` includes
+24 vLLM Python files for GLM metadata reuse, dense-kernel integration,
+the NVFP4 proposal head, deferred-weight ownership, and independent draft
+and rejection-sampling randomness. The adjacent patch provides a readable diff;
+`source-lock.json` records the base, donor revisions, and exact file hashes.
+These files derive from Local Inference Lab's vLLM fork at `3512b066` and
+`a8c796f3`, loader correction `17e341b9`, and independent-RNG backport
+`44e6766e` from [PR 653](https://github.com/local-inference-lab/vllm/pull/653),
+under Apache-2.0 with their contributor notices retained.
 
 The unified diffs under `runtime/deepseek0731-gb10/patches/` contain context
 and removed lines from vLLM, pinned to the source revision recorded by that
@@ -245,8 +255,10 @@ artifacts that are not distributed in this repository:
   repository notices.
 
 SparkRing records these identities and validates compatible image content; it
-does not redistribute the model weights, vLLM source, or B12X model-kernel
-package. The selected B12X communication source in Section 11 is included. Operators
+does not include model weights or the complete vLLM/B12X source trees in this
+Git repository. The selected vLLM files in Section 3 and B12X communication
+source in Section 11 are included. Published runtime images contain the
+pinned vLLM and B12X packages under their respective licenses. Operators
 must obtain each artifact under its own terms. The exact operator-image
 composition is in `runtime/glm53-flash-jj-r8-gb10/pins.json` and
 `runtime/glm53-flash-jj-r8-gb10/glm53-dcp4-sircl-public-image-receipt.json`.
@@ -287,3 +299,16 @@ The native-MTP3 profile references
 Operators must obtain and use that checkpoint under its own license and
 notices. Native MTP uses its included prediction layer and does not require
 an external DFlash checkpoint.
+
+## 12. B12X selector source and MoE scale sharing
+
+The native-MTP3 compute package downloads B12X revision
+`ef308bac0f3b3eb8fea63e4013afc0c2ea1c6301`, including its shared native NVFP4
+scales for A4/A16 MoE paths. Three selector Python files are included in
+`runtime/glm53-spark-mtp3-mesh/compute/b12x-selector-files.tar.gz` from
+[B12X PR 316](https://github.com/local-inference-lab/b12x/pull/316), revision
+`9ac142824b4edb750892a0fb63d914230086495d`. They implement the top-k-512
+candidate buffer, exact overflow handling, and omission of unused terminal
+scores. These files are licensed under Apache-2.0; source notices are retained
+and the downloaded B12X archive supplies the license. The compute source lock
+binds the source archive and each base/result file hash.