Publish the integrated native-MTP3 cache and checkpoint profile

.gitattributes

@@ -2,6 +2,8 @@
 *.patch text eol=lf
 runtime/** text eol=lf
 runtime/sparkring/*.gz -text
+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

@@ -29,12 +29,12 @@
 
 | Model / predictor | Layout | Context | Sequences | Batch | Guide |
 |---|---|---:|---:|---:|---|
-| **GLM-5.3 Flash NVFP4-Spark · MTP3 mesh** | TP4/DCP4 | 1M | 16 | 8,192 | [Quickstart](docs/GLM53_SPARK_MTP3_MESH_QUICKSTART.md) |
+| **GLM-5.3 Flash NVFP4-Spark · MTP3 cache/checkpoint mesh** | TP4/DCP4 | 1M | 16 | 8,192 | [Quickstart](docs/GLM53_MTP3_CACHE_CHECKPOINTS_QUICKSTART.md) |
 | GLM-5.3 Flash NVFP4 · BF16 DFlash2 | TP4/DCP4; DCP1/2 | 1M | 16 | 8,192 | [Quickstart](docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md) |
 | GLM-5.2 EXL3 3.5-bpw | TP4/DCP4 | 1M | 16 | 4,096 | [Quickstart](docs/GLM52_35BPW_QUICKSTART.md) |
 | DeepSeek-V4-Flash-0731 | TP4/DCP1 | 1M | 32 | 4,096 | [Quickstart](docs/DEEPSEEK_V4_FLASH_QUICKSTART.md) |
 | Qwen3.8-27B EXL3 K5/K6 | TP4/DCP1 | 1M | 64 | 8,192 | [Quickstart](docs/QWEN38_27B_EXL3_K5K6_QUICKSTART.md) |
 * The GLM5.3 Flash MTP3 profile uses the new virtual meshing feature.
 * Requires:[managed-mesh setup](runtime/glm53-spark-mtp3-mesh/MANAGED_MESH.md). 
 * DFlash2 profiles will be replaced by native-MTP. Also avoids: [separate CC BY-NC-ND 4.0 terms](https://huggingface.co/incoai/GLM-5.3-Flash-DFlash2#license).
 
@@ -45,7 +45,7 @@
 | **GLM-5.3 Flash NVFP4-Spark · native MTP3** | TP2/DCP1 | 512K | 8 | 8,192 | [Quickstart](docs/GLM53_FLASH_SPARK_TP2_EXPERIMENTAL_QUICKSTART.md) |
 | DeepSeek-V4-Flash-0731 | TP2/DCP1 | 1M | 32 | 4,096 | [Quickstart](docs/DEEPSEEK_V4_FLASH_QUICKSTART.md) |
 | Qwen3.8-27B EXL3 K5/K6 | TP2/DCP1 | 1M | 32 | 8,192 | [Quickstart](docs/QWEN38_27B_EXL3_K5K6_PAIR_QUICKSTART.md) |
 * GLM5.3 pair is **research-only**, uses 5 GiB KV per rank, and has a
 [known video-color issue](https://github.com/FujitsuPolycom/sparkring/issues/229). In testing. 
 See the [profile index](docs/profiles/README.md) for evidence scopes and
 [SparkCache compositions](recipes/sparkcache/README.md) for persistent-cache

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.

docs/GLM53_JJ_R8_GB10_SPARKCACHE_TP4_QUICKSTART.md

@@ -305,6 +305,11 @@ SPARKCACHE_ACCESS_MODE=read-write   # restore existing entries and publish new o
 SPARKCACHE_ACCESS_MODE=restore-only # restore existing entries; never capture new prompts
 ```
 
+Keep `SPARKCACHE_ASYNC_PAGE_CAPTURE=auto` from the environment template to
+disable capture automatically in restore-only mode or when SparkCache is off.
+An explicit `1` with either setting is rejected. Explicit `0` in read-write
+mode selects synchronous publication.
+
 The GLM-5.3 profile retains a verified shared GPU prefix for up to five
 minutes so one restore can serve an extended request queue:
 
@@ -360,20 +365,34 @@ and `IMAGE_ID`; keep the page-tail settings unchanged.
 The recommended DCP4 profile enables bounded asynchronous page capture:
 
 ```bash
-SPARKCACHE_ASYNC_PAGE_CAPTURE=1
+SPARKCACHE_ASYNC_PAGE_CAPTURE=auto
 SPARKCACHE_ASYNC_CAPTURE_SLOT_BYTES=auto
 SPARKCACHE_ASYNC_CAPTURE_SLOT_COUNT=2
 ```
 
 The `auto` slot policy selects 8 GiB for DCP1, 5 GiB for DCP2, or 3 GiB for
 DCP4. Two capture slots let the background publisher consume one completed
 capture while a later capture uses the other. Restore separately overlaps
-bounded NVMe reads and CUDA placement through two 256 MiB mapped arenas. More
-restore arenas are not part of this profile because measured arena waits did
-not justify the additional unified-memory pressure. DCP1 and DCP2 page-tail
+bounded NVMe reads and CUDA placement through two 256 MiB mapped arenas per
+load lane. Eight lanes reserve 4 GiB per rank; adding the DCP4 capture slots
+gives 10 GiB per rank (40 GiB across TP4) of configured payload capacity,
+before KV, models, metadata, retained objects and transport. Restore-only
+omits the 6 GiB capture ring. DCP1 and DCP2 page-tail
 capture have no matching live record; use complete snapshots or test those
 layouts separately.
 
+Print the resolved memory plan without host or GPU access before launching:
+
+```bash
+SPARKRING_PRINT_MEMORY_PLAN=1 bash runtime/glm53-flash-jj-r8-gb10/launch-rank.sh \
+  0 /path/to/rank.env
+```
+
+Set `SPARKCACHE_BUFFER_BUDGET_BYTES` in that configuration to enforce a
+per-rank restore-plus-capture payload ceiling; zero disables it. The report
+shows KV separately and names excluded memory. It does not predict total
+resident memory or replace a hardware capacity test.
+
 The environment template enables `DFLASH_WARMUP=1`. Rank 0 waits for the API,
 then exercises every concurrency from C1 through C16 and scheduled prompt
 spans covering DFlash's Triton block-size specializations. DFlash depth seven

docs/GLM53_MTP3_CACHE_CHECKPOINTS_QUICKSTART.md

@@ -0,0 +1,98 @@
+# GLM-5.3 native MTP3 with verified caching and recurrent checkpoints
+
+Status: **research-only**. The image build and source checks are implemented.
+The published image's vLLM, B12X, transport, and warmup files match the deployed
+composition in 5,308 byte comparisons. It incorporates SparkCache main at the
+revision below. Source equivalence is not a serving soak of this rebuilt image.
+
+## Prerequisites
+
+Use four NVIDIA Sparks with the [managed mesh host setup](GLM53_SPARK_MESH_HOST_SETUP.md).
+The model uses TP4/DCP4 and native MTP depth three. Keep the reviewed fabric,
+driver, GID and interface settings; this guide does not reconfigure NICs.
+Follow [host prerequisites](PREREQUISITES.md#four-spark-managed-hardware-forwarded-mesh),
+including optional reboot preparation after large GPU workloads. Stop model
+workloads through the managed lifecycle before replacing containers or services.
+
+Download the target `local-inference-lab/GLM-5.3-Flash-NVFP4-Spark` at revision
+`df116c4fb16b1d37ae43d2cfd624de26ffbc832e`. No external draft is required.
+
+## Pull and verify the image
+
+Use the same immutable image on all four ranks:
+
+```bash
+image='ghcr.io/fujitsupolycom/sparkring-glm53-sparkcache@sha256:11a556a54041fd823d152a7f051ac4f7c617dc539030df26e93008392fee0746'
+image_id='sha256:6921a6c163ea40b603e19a0332330efe3dbccbf4dce9f6cbbf6b756c9231835a'
+docker pull "$image"
+test "$(docker image inspect "$image" --format '{{.Id}}')" = "$image_id"
+docker run --rm --network none --entrypoint python3 "$image" \
+  -S -B /opt/sparkring/bin/verify-performance.py
+```
+
+The [public image contract](../runtime/glm53-spark-mtp3-mesh/performance/public-image.json)
+is the renderer and installer input for this composition. Do not substitute the
+compute parent's `image-receipt.json`. Source builds use
+[the performance build instructions](../runtime/glm53-spark-mtp3-mesh/performance/README.md).
+
+## Extract and render
+
+Choose unused artifact and export-container names. The following commands do
+not start a model. Create `/srv/sparkring/artifacts` with operator permissions:
+
+```bash
+docker create --name mtp3-artifact-export "$image"
+docker cp mtp3-artifact-export:/opt/spark-sircl /srv/sparkring/artifacts/cache-checkpoints-bundle
+docker cp mtp3-artifact-export:/opt/sparkring/bin/mlx5-rdma-tx-marker /srv/sparkring/artifacts/mlx5-rdma-tx-marker
+docker rm mtp3-artifact-export
+cp runtime/glm53-spark-mtp3-mesh/performance/public-image.json /srv/sparkring/verified-image-receipt.json
+```
+
+Edit the private site and fabric inputs described in the
+[managed deployment guide](../runtime/glm53-spark-mtp3-mesh/MANAGED_MESH.md).
+Set the site's bundle root to the extracted bundle and its marker path to the
+extracted marker. The marker SHA-256 is
+`2828c07e4255c4962c77425be2c88969e7eb7dd4b1bf9e36485bc705bb5d6d64`.
+Use a distinct container prefix and cache directory; do not rename entries from
+another cache namespace into this one.
+
+```bash
+python runtime/glm53-spark-mtp3-mesh/profile.py render \
+  --site /srv/sparkring/site/mtp3-mesh.json \
+  --bundle /srv/sparkring/artifacts/cache-checkpoints-bundle \
+  --image-receipt /srv/sparkring/verified-image-receipt.json \
+  --output build/mtp3-cache-checkpoints-launch
+```
+
+The renderer selects the matching image, transport manifest, optimized native
+placement digest, and cache namespace. Review the generated rank environments
+and plan, then follow the managed guide's create-only container, installation,
+and coordinated startup steps with this launch directory and verified receipt.
+Do not use direct `docker start` to bypass the memory and four-rank gates.
+
+## Runtime behavior and limits
+
+SparkCache revision `48bbd2be4a7b972e56632a2d7b934bac5460f272` provides bounded
+restoration, tiled CUDA placement, authenticated history reconstruction,
+publication dependency protection, and publication-backlog gauges. Periodic
+full captures remain off by default. They trade more writes for shorter
+history processing and require a separately selected workload configuration.
+
+The profile retains 24 GiB KV capacity per rank, an 8,192-token batch budget,
+40 GiB persistent-cache maximum and 32 GiB low watermark. Those capacity
+defaults are not evidence of a 40 GiB multimodal stress test. The recorded
+cache-pressure soak used 2 GiB per rank and an opt-in full-capture cadence.
+
+The image includes explicit two-checkpoint prefill, complete convolution-state
+exports, stream-ordered transport, CTA publication synchronization, delayed
+proxy-thread startup, and linked payload/doorbell posts. Checkpoint source and
+ownership hashes are preserved by the build rather than inferred from labels.
+
+GLM-5.3 supports `reasoning_effort=low`, `high`, or `max`. Explicit thinking-off
+chat requests return HTTP 400; warmup uses supported reasoning settings. This
+does not implement reasoning-free generation. Responses API behavior is not
+qualified by the chat-completion checks.
+
+Run bounded semantic, repeated-prefix, multimodal, and idle-probe checks after
+deployment. Neither source equivalence nor the parent's measurements qualify
+unattended availability, every cache boundary, or a universal speedup.