Pin the live-tensor-compatible B12X runtime

.github/workflows/ci.yml

@@ -53,7 +53,7 @@ jobs:
             --index-url https://download.pytorch.org/whl/cpu \
             "torch==${TORCH_VERSION}"
       - name: Test maintained Python trees
-        run: python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
+        run: python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-b12x-kda-adaptive-mtp runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
 
   docs-links:
     name: docs links

AGENTS.md

@@ -104,7 +104,7 @@ test imports torch:
 python -m pip install -r requirements-dev.txt
 python -m pip install --index-url https://download.pytorch.org/whl/cpu "torch==2.11.0"
 ruff check --select E,F,W --ignore E501 spark_transport runtime scripts performance
-python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
+python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-b12x-kda-adaptive-mtp runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
 ```
 
 The test suite is CPU-only contract coverage. It does not validate CUDA,

docs/GLM53_B12X_KDA_ADAPTIVE_MTP_SPARKCACHE_TP4_QUICKSTART.md

@@ -1,9 +1,17 @@
 # Serve GLM-5.3 with adaptive MTP, live-tensor B12X KDA, and SparkCache
 
 Status: **implemented, not qualified**. This guide builds vLLM commit
-`0b67266a0f37d6146a8403fb8482403c62f412d5` and the SparkCache overlay from
-commit `20838ace3ebda570ca039cb7f1976c29da554b39` for four DGX Spark systems at
-TP4/DCP1.
+`0b67266a0f37d6146a8403fb8482403c62f412d5`, B12X commit
+`b1d541f9e71a35f030d45fae437630fff7507c2a` with Git tree
+`c69cdec1c59a08e8e0e549f930fa8abcfb5134ae`, and the SparkCache overlay
+from commit `20838ace3ebda570ca039cb7f1976c29da554b39` for four DGX Spark systems
+at TP4/DCP1.
+
+The B12X implementation accepts `kda_metadata_validation="trusted"` and binds
+each request's live projection, metadata, and output tensors at their actual
+capacities. B12X commit `2fcf23a0ce269be27b2e03fece73d46e90e6aeea`
+is **unsupported** with this vLLM revision because it lacks both parts of that
+interface.
 
 The serving profile uses embedded MTP with maximum depth five, initial depth
 three, and a 32-step acceptance window. Fastsafetensors uses queue size one.
@@ -55,9 +63,11 @@ test "${#native_sha256}" -eq 64
 ```
 
 The runtime builder verifies the complete first-parent vLLM history from
-`da4d7be` through adaptive MTP and the three live-tensor B12X KDA commits. The
-SparkCache build verifies LF Linux preimages, four exact patches, and eleven
-postimage source files.
+`da4d7be` through adaptive MTP and the three vLLM live-tensor KDA commits. It
+also verifies the B12X first-parent history from `2fcf23a` through `b1d541f9`,
+the exact KDA source hashes, trusted metadata selection, and request-sized
+live-tensor binding. The SparkCache build verifies LF Linux preimages, four
+exact patches, and eleven postimage source files.
 
 ## Resolve the TP4 profile
 
@@ -125,10 +135,13 @@ NCCL error, or traceback.
 
 The overlay does not change SparkCache wire fields, digest salts, 256-token
 geometry, or stored object schemas. Its embedded-MTP digest is SHA-256 over
-`glm53-embedded-mtp-runtime-v1`, the target identity, the full vLLM commit,
-maximum depth five, and `adaptive:3:32`, separated by zero bytes.
-
-Including the vLLM revision gives this runtime a distinct draft-state cache
-identity from the e105 adaptive-MTP profile. Stored entries therefore
-recompute instead of crossing the KDA source boundary without byte-equivalence
-evidence.
+`glm53-embedded-mtp-vllm-b12x-runtime-v1`, the target identity, the full vLLM
+commit, the full B12X commit, maximum depth five, and `adaptive:3:32`, separated
+by zero bytes. The resulting digest is
+`2761488b43742f849e2cb7ac4977dc28d771faf5afef1951044327948d6ad71e`.
+
+Including both implementation revisions gives this runtime a distinct
+draft-state cache identity from the e105 adaptive-MTP profile and from any
+0b67266 runtime paired with another B12X source. Stored entries therefore miss
+and recompute instead of crossing a KDA implementation boundary without
+byte-equivalence evidence.

runtime/README.md

@@ -72,7 +72,7 @@ allowlist and validating the generated manifest through the builder.
 The builder test suites cover the GLM, DeepSeek, and Qwen runtime contracts:
 
 ```bash
-python -m pytest runtime/exl3-r7 runtime/glm53-flash runtime/deepseek0731-gb10 runtime/qwen38 -q
+python -m pytest runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-b12x-kda-adaptive-mtp runtime/deepseek0731-gb10 runtime/qwen38 -q
 ```
 
 [`runtime/glm53-flash/`](glm53-flash/README.md) builds the source-pinned

runtime/glm53-flash-b12x-kda-adaptive-mtp/Containerfile

@@ -5,8 +5,10 @@ ARG ARM_BUILDER
 
 FROM ${ARM_BUILDER} AS b12x-wheel
 ARG B12X_COMMIT
+ARG B12X_TREE
 COPY bundle/sources/b12x /build/b12x
 RUN test "$(git -C /build/b12x rev-parse HEAD)" = "${B12X_COMMIT}" \
+ && test "$(git -C /build/b12x rev-parse 'HEAD^{tree}')" = "${B12X_TREE}" \
  && /opt/python/cp312-cp312/bin/python -m pip wheel \
       --no-deps --wheel-dir /out/wheels /build/b12x
 
@@ -27,6 +29,8 @@ RUN test "$(git -C /build/nccl rev-parse HEAD)" = "${NCCL_COMMIT}" \
 FROM ${VLLM_BASE}
 ARG VLLM_COMMIT
 ARG B12X_COMMIT
+ARG B12X_TREE
+ARG B12X_KDA_CONTRACT
 ARG NCCL_COMMIT
 ARG NCCL_PATCHED_TREE
 ARG NCCL_PATCH_SHA256
@@ -35,7 +39,7 @@ ARG SPARKRING_REVISION
 
 COPY --from=b12x-wheel /out/wheels/ /opt/sparkring/wheelhouse/
 RUN uv pip install --system --no-deps /opt/sparkring/wheelhouse/b12x-*.whl \
- && python3 -c "import b12x; from vllm.utils.b12x import get_b12x_sparse_mla, get_b12x_dsa_indexer, get_b12x_gdn_decode, get_b12x_mhc; assert get_b12x_sparse_mla(); assert get_b12x_dsa_indexer(); assert get_b12x_gdn_decode(); assert get_b12x_mhc()"
+ && python3 -c "import b12x, inspect; from vllm.utils.b12x import get_b12x_sparse_mla, get_b12x_dsa_indexer, get_b12x_gdn_decode, get_b12x_mhc; gdn = get_b12x_gdn_decode(); assert get_b12x_sparse_mla(); assert get_b12x_dsa_indexer(); assert gdn; assert get_b12x_mhc(); assert gdn.Caps.__dataclass_fields__['kda_metadata_validation'].default == 'transactional'; bind_source = inspect.getsource(gdn.bind_kda); run_source = inspect.getsource(gdn.run_kda); assert 'mixed_qkv.shape[0]' in bind_source; assert 'state_indices.shape[0]' in bind_source; assert 'caps.kda_metadata_validation == \"transactional\"' in run_source"
 
 COPY --from=nccl-build /out/nccl/ /opt/sparkring/nccl/
 RUN cd /opt/sparkring/nccl \
@@ -77,6 +81,8 @@ LABEL org.opencontainers.image.title="SparkRing GLM-5.3 Flash ARM64 runtime" \
       org.jovian.architecture="linux-arm64-sm121" \
       org.jovian.vllm.commit="${VLLM_COMMIT}" \
       org.jovian.b12x.commit="${B12X_COMMIT}" \
+      org.jovian.b12x.tree="${B12X_TREE}" \
+      org.jovian.b12x.kda-contract="${B12X_KDA_CONTRACT}" \
       org.jovian.transport="sparkring-nccl-2.30.7-source-built" \
       org.sparkring.nccl.commit="${NCCL_COMMIT}" \
       org.sparkring.nccl.patched-tree="${NCCL_PATCHED_TREE}" \

runtime/glm53-flash-b12x-kda-adaptive-mtp/README.md

@@ -2,16 +2,26 @@
 
 Status: **implemented**. The builder pins
 `local-inference-lab/vllm@0b67266a0f37d6146a8403fb8482403c62f412d5`
-and Git tree `ba9484ccb33aa56e90ff2f447f15ca9b9da97639`. Four-rank
-TP4/DCP1 serving remains **unqualified** until an immutable image digest has a
-live qualification receipt.
+at Git tree `ba9484ccb33aa56e90ff2f447f15ca9b9da97639` and
+`local-inference-lab/b12x@b1d541f9e71a35f030d45fae437630fff7507c2a`
+at Git tree `c69cdec1c59a08e8e0e549f930fa8abcfb5134ae`. Four-rank TP4/DCP1
+serving remains **unqualified** until an immutable image digest has a live
+qualification receipt.
 
 The pinned vLLM history contains the complete GLM-5.3 source runtime from
 `da4d7be6c97434f6942292ed8abbf4b32dc44355` through acceptance-based
 adaptive MTP at `e10536aadf02a18fccddda7ec939c33147e8b0b3`, followed by three
 commits that bind B12X KDA metadata once and operate on live layer tensors.
 `pins.json` records and verifies the exact first-parent sequence.
 
+The B12X pin accepts `kda_metadata_validation="trusted"` and derives tensor
+capacity from each live request's projection and recurrent-state metadata.
+B12X commit `2fcf23a0ce269be27b2e03fece73d46e90e6aeea` is **unsupported** for this
+runtime: its `Caps` interface has no trusted-metadata selection, and its
+`bind_kda` interface requires plan-sized tensors. The source preparer rejects
+that byte-exact implementation and any source that lacks the required call
+contract.
+
 The runtime also pins B12X, InstantTensor, CUDA, and SparkRing's source-built
 NCCL transport. The fastsafetensors TP4 profile uses loader queue size one.
 The vLLM implementation selects `nogds=True` when tensor parallelism exceeds
@@ -26,15 +36,17 @@ BUILD_RECEIPT="$PWD/glm53-b12x-kda-adaptive-mtp-runtime-receipt.json" \
 bash runtime/glm53-flash-b12x-kda-adaptive-mtp/build-image.sh
 ```
 
-The builder verifies commits, Git trees, the complete vLLM lineage, source
-licenses, patched NCCL bytes, output labels, and required Python imports. Its
-receipt proves image construction only. Startup, semantic generation,
-SparkCache restore, shared-prefix concurrency, and fatal-log checks require a
-separate four-rank receipt.
+The builder verifies commits, Git trees, the complete vLLM and B12X lineages,
+the trusted-metadata request-sized KDA call contract, source licenses, patched
+NCCL bytes, output labels, and required Python imports. Its receipt proves
+image construction only. Startup, semantic generation, SparkCache restore,
+shared-prefix concurrency, and fatal-log checks require a separate four-rank
+receipt.
 
 The matching SparkCache overlay is pinned to
 `FujitsuPolycom/sparkcache@20838ace3ebda570ca039cb7f1976c29da554b39`.
 Its Linux-byte-exact vLLM contract is
 `vllm-kv-block-lease-contract-glm53-b12x-kda-adaptive-mtp.json`. The
-runtime-bound embedded-MTP identity prevents this profile from reusing e105
-adaptive-MTP entries until byte-equivalence across the KDA revisions is proven.
+embedded-MTP identity binds both the vLLM and B12X revisions. Entries produced
+by another KDA implementation therefore miss and recompute instead of crossing
+an unproven numerical boundary.

runtime/glm53-flash-b12x-kda-adaptive-mtp/build-image.sh

@@ -40,6 +40,8 @@ arm_builder="$(read_pin public_image_build.base_images.arm_builder.reference)"
 cuda_runtime="$(read_pin public_image_build.base_images.cuda_runtime.reference)"
 vllm_commit="$(read_pin public_image_build.sources.vllm.commit)"
 b12x_commit="$(read_pin public_image_build.sources.b12x.commit)"
+b12x_tree="$(read_pin public_image_build.sources.b12x.tree)"
+b12x_kda_contract="$(read_pin public_image_build.vllm_b12x_kda.schema)"
 nccl_commit="$(read_pin public_image_build.sources.nccl.commit)"
 nccl_patched_tree="$(read_pin public_image_build.sources.nccl.patched_tree)"
 nccl_patch_sha256="$(read_pin public_image_build.sources.nccl.patches.0.sha256)"
@@ -106,6 +108,8 @@ vllm_image="sparkring-glm53-vllm:${vllm_commit:0:12}-arm64"
   --build-arg "ARM_BUILDER=${arm_builder}" \
   --build-arg "VLLM_COMMIT=${vllm_commit}" \
   --build-arg "B12X_COMMIT=${b12x_commit}" \
+  --build-arg "B12X_TREE=${b12x_tree}" \
+  --build-arg "B12X_KDA_CONTRACT=${b12x_kda_contract}" \
   --build-arg "NCCL_COMMIT=${nccl_commit}" \
   --build-arg "NCCL_PATCHED_TREE=${nccl_patched_tree}" \
   --build-arg "NCCL_PATCH_SHA256=${nccl_patch_sha256}" \

runtime/glm53-flash-b12x-kda-adaptive-mtp/pins.json

@@ -49,8 +49,29 @@
       },
       "b12x": {
         "repository": "https://github.com/local-inference-lab/b12x.git",
-        "commit": "2fcf23a0ce269be27b2e03fece73d46e90e6aeea",
-        "tree": "58a046fc8faa747346f40f87166cda7e0f67ff47",
+        "commit": "b1d541f9e71a35f030d45fae437630fff7507c2a",
+        "tree": "c69cdec1c59a08e8e0e549f930fa8abcfb5134ae",
+        "source_lineage": {
+          "base_commit": "2fcf23a0ce269be27b2e03fece73d46e90e6aeea",
+          "included_commits": [
+            "f1ddd12fe51667ade19a6fb6b460e338f0e94461",
+            "d56c1163b6e019d828ed24f135c2efd05fdca6ea",
+            "f2fd5113db6344a117074424e5461329ba33704f",
+            "11814a27cabc3501f3ab8d20f159a80f6d29d0af",
+            "ff9abf8a08fd4af0d3f7b3bab69d500e924bab2e",
+            "e8905c1afaedb1b905dc8dd80537dbc25d35453b",
+            "5c20a7f3190f92fabd290ddbd2a77390fe2b0f8f",
+            "501ffb9a9b118009b8ea2ea20b8d43f57ba4f221",
+            "98149b404f59498324d3193ccbd1a49aa8df21a8",
+            "cd9c3431f760f965fcf5f268b6a8d6ef1de9b6db",
+            "80d68e5702e7c9e99ca37988f87c964021d7e1b4",
+            "4e65ed4d7a57b708bc213580c19a4e9cf0a9ca82",
+            "12aea7d96928f540a64259e4e24ef7688093b515",
+            "f1a37c87a1b81cfe7b302b0a8c148ca4302ca8dc",
+            "00f7d4be8c080e6faf359c089d1439839565b87e",
+            "b1d541f9e71a35f030d45fae437630fff7507c2a"
+          ]
+        },
         "license": "Apache-2.0"
       },
       "nccl": {
@@ -69,6 +90,29 @@
         ]
       }
     },
+    "vllm_b12x_kda": {
+      "schema": "sparkring-vllm-b12x-live-tensor-kda/v1",
+      "status": "implemented",
+      "metadata_validation": "trusted",
+      "tensor_binding": "request-sized-live-tensors",
+      "vllm_source": {
+        "path": "vllm/model_executor/layers/mamba/gdn/kimi_gdn_linear_attn.py",
+        "sha256": "8bf8bc579dd4a80224dc1633e7513f2a0c58e07db72a736c7e41d28d3c35f3b9"
+      },
+      "b12x_source": {
+        "path": "b12x/sequence/gdn_decode/_impl.py",
+        "sha256": "a2c81a486eb4d86a59f39e4f03855382583d39cdadd5fc8fd1c685ba28b1b56e"
+      },
+      "unsupported_b12x_sources": [
+        {
+          "status": "unsupported",
+          "commit": "2fcf23a0ce269be27b2e03fece73d46e90e6aeea",
+          "tree": "58a046fc8faa747346f40f87166cda7e0f67ff47",
+          "source_sha256": "b1f072405ad2f3bba83e720419a59ae791e1269dfccc9fc4a279889f8bd07d6e",
+          "limitation": "Caps does not accept kda_metadata_validation, and bind_kda requires plan-sized tensors instead of request-sized live tensors."
+        }
+      ]
+    },
     "instanttensor": {
       "version": "0.1.9",
       "sdist_url": "https://files.pythonhosted.org/packages/37/69/a4dc4e0f0018a0e558b716e90bcb19cb6a8c7506f68e89a1493608cf5e62/instanttensor-0.1.9.tar.gz",