RC4.1: faster prose and code decode, bit-exact with RC2; long GLM tool calls fixed

[data-angel]

RC4.1 hotfix (current pin). The engine is Enntity/atlas sparkglm/atlas-20261012-rc4.1-layered @ 301479a2 (tree ac5a080b). The image is ghcr.io/enntity/atlas-sparkglm:cfb463a952b5 (sha256:0ddea7cdb372e4407db63dc4f2b5b1a08b4576b8223ded6a4fca3463439ccfb2), install tree cfb463a952b517b87baf302b5b3b279a115c5626. It is staged on both pair hosts and not yet deployed.

• The bug. The envelope-stuck guard (MAX_TOOL_BODY_TOKENS = 1024) recognized parameter values only by Qwen's <parameter=…> token ids. Every GLM tool call longer than about 1,024 tokens therefore ended with finish_reason "length" and no tool call. The bug has been present since before RC2.

• The fix (ade7b012 on fix/glm-tool-value-markers, on top of upstream/glm53-flash 0b515eec): the engine resolves GLM's single-token <arg_value> / </arg_value> at startup and tracks value bodies by them. Sampling is unchanged.

• Layering. The seven SparkGLM-only layer commits are re-applied on top of the fix. The first of them (661a26a0) now declares the KV-shard flag its native bridge reads, which upstream 7e13c1bb removed.

• Against RC4's measured tree (5c523903 / ae5e79f8), the pin adds exactly two things, and neither touches the decode path:

  • the fix, which is spark-server only;
  • the four CI-only upstream commits listed below.

  The speed and exactness receipts in results/2026-10-12-rc4/ were measured on RC4. RESULT.md has an RC4.1 section. Live gate results will be added after the deploy.

• Profiles are unchanged from RC4.

---

RC4 moves the engine to Enntity/atlas sparkglm/atlas-20261012-rc4-layered @ 5c523903. That is the upstream candidate upstream/glm53-flash-20261012-rc4 @ ad047638 plus the unchanged SparkGLM-only layer. Its tree ae5e79f8 is identical to exp/prose-rc4 @ c6089633, the build measured on the pair.

The image is ghcr.io/enntity/atlas-sparkglm:846ee9e1e501 (sha256:88461891963193b8e2abf0cb1cd603bf31fd9e3a8e1422074d10dd46f886485f). The install tree is 846ee9e1e501eaf7afb71211aa2473d9610b2344. It is deployed on the pair through LLooM.

Install. All three profiles add ten engine switches and one profile switch:

• ATLAS_RDMA_ONESHOT_SM_STAGE=1
• ATLAS_GLM_MOE_ROUTE_FUSE=1
• ATLAS_GLM_DENSE_TP_SPLIT=1
• ATLAS_GLM_KDA_COMMIT_CTAS=96
• ATLAS_GLM_DRAFT_TP_TOPK=1
• ATLAS_DFLASH_ASYNC_UPLOADS=1
• ATLAS_GLM_MLA_FUSE=1
• ATLAS_GLM_MXFP8_TILED=1
• ATLAS_DFLASH2_EDGE=0.6
• ATLAS_DFLASH_CONV_TWIN=nvfp4
• ATLAS_DFLASH_CONF_PROBE=256 (the profile switch)

With SPARKGLM_KV_SHARD=1, the 4x1m environment equals the tested dev profile exactly. Only 4x1m was measured. install/atlas-source.json pins 5c523903 and records the layers.

Results (results/2026-10-12-rc4/; raw receipts with SHA256SUMS). All results are bit-exact with RC2:

C1 decode	RC4	RC2
Prose, 12 prompts × 256 tokens	45.57 tok/s	42.82
Strict prose bursts, 8 × 384 tokens	47.9	44.8
Code, 12 prompts × 256 tokens	64.91	61.38

Exactness evidence:

• the logprob hash 246cd1331b41 is unchanged;
• 524/524 probe outputs are byte-identical to RC2's;
• C4 output equals C1;
• the four-session 204K fill on the deployed recipe scored 12/12.

README numbers changed:

• The single-stream decode row is now 45.6 / 64.9 tok/s (it was 42.9 / 61.7). Its reference cell is now RC2's 42.8 / 61.4, and the TensorFold clause in that cell is gone.
• A link to this release's receipts was added.
• No other rows were re-run.
• The README graphic and cards belong to the README/card agent and were not regenerated. Several other rows still carry the existing competitor references.

Later upstream fixes, not in this release. The upstream branch upstream/glm53-flash-20261012-rc4 (and upstream/glm53-flash, fast-forwarded to it) now carries four CI-only fix commits on top of ad047638. None of them changes GB10 runtime behavior. The only runtime-source edit removes a variable nothing reads on that branch. They are:

• 7e13c1bb drops the unused sharded flag in paged_glm.rs. Only the SparkGLM layer reads it, so the next re-layer must declare it inside the layer commit.
• bc9d0de4 replaces the mxfp8_tile16 intra-doc link with plain code.
• a6f35bf7 mirrors the DF2_SEL_EDGE twin into the strix-hip selector copy.
• 0b515eec applies rustfmt to canonical_verify_gpu_tests.rs.

RC4 pinned the tested tree at 5c523903. RC4.1 (above) builds on upstream/glm53-flash with these four commits included.

After merge: in LLooM atlas-sparkglm-rc4 (Enntity/lloom#64), replace the SparkGLM revision 35dc1b41c92351d09d47af64e69b44b3393e2334 (and its short form 35dc1b41) with the merge commit. It appears in the pin, recipe, docs and test files that tools/lloom_pin_rc41.py rewrites. The tree, tag and digest stay the same.

🤖 Generated with Claude Code