Skip to content

GLM-5.3-Flash native NCCL P2P transport A/B evidence

This directory contains the sanitized public evidence for 2026-09-09-glm53-native-nccl-p2p.

Fixed campaign boundary

  • Model: ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO at c3cbb9891b67c741bcbf6b176dd7af9265b069db
  • Runtime: ghcr.io/ormandj/sglang-glm53-flash-sm120@sha256:0c0637959c3931829f05154087bbefd2c50003fb9b2010200ce0ec82f4d71a53
  • Measured configuration: native Linux kernel 7.0.0-31, driver 595.91.07, two RTX PRO 6000 Blackwell Max-Q GPUs, TP=2, 393,216 tokens, C1, and adaptive MTP fixed.
  • Only intended A/B delta: NCCL_P2P_DISABLE=1 versus NCCL_P2P_DISABLE=0; NCCL_CUMEM_ENABLE=0, IOMMU-off state, and ACS state are fixed.
  • Capacity protocol: 4K/120K target contexts at n=12/3 for comparative performance, with unique prompt cache mode, request canaries, strict 32-code-word output, and a 512-token maximum. The earlier 128-word scout and 380K n=1 capacity attempt remain retained negative evidence; the latter reached 304,588 actual prompt tokens but returned 33 rather than 32 code words and is not performance-eligible. Matched 380K target retrieval and long-tool assertions are separate correctness gates.

Finalization and validation

The artifact manifest retains all ten required roles. Native request artifacts remain authoritative. Public files redact private identifiers while preserving native schemas and a redaction record. The graph and manifest helpers each produced byte-identical results across two runs. The 19 focused publication tests, tracked Markdown link check and strict MkDocs build passed. The strict build used the repository's docs requirements in an isolated tool environment.

The authorized native GLM recipe retains P2P enabled. See the decision summary, publication summary, comparison, and chart. The two 380K strict-output failures and the raw diagnostic marker failure remain retained limitations.