b70-optimization-lab

Laguna source and runtime reconstruction

This packet separates three claims that must not be blurred:

1. Sealed-evidence audit is portable and performs no GPU work. The tracked record evidence, source bundles, model manifest, token/text oracles, accounting, graph evidence, cleanup, and LocalMaxxing receipt can all be verified from Git.
2. Artifact-exact lab replay uses the originating host’s byte-identical virtual environment and native libraries. run.sh --preflight hashes every direct and transitively loaded native object and checks its actual loader origin before a service can start.
3. Source-equivalent rebuild restores all source and model revisions but is a new environment. Compiler and linker output is not promised to be byte-identical. It must pass the same 13/13 token/text, cache-zero, 146/145 topology, treatment, cleanup, and explicit performance-floor gates before it is called a reproduction.

The historical build was a mixed native stack assembled during an optimization campaign. The original single monolithic build command was not sealed, so this file does not invent one.

Restore all source provenance

cd /home/steve/llm-optimizations
repro/laguna-s-2.1-int4-b70-102tps-20260726/restore-sources.sh \
  /mnt/fast-ai/laguna-repro-sources

The command restores five clean worktrees:

tree	commit	role in the measured native stack
vLLM record	e596ef1543466ae1a05e5bb8091f58872e2b18ba	Python/runtime record source
kernel record	6f9dd3c3a7b1b677a992ca4f431a968408f9c816	final router/workspace source and _moe_C
attention runtime	906190641d708b8028018c5dde653e265c835348	_vllm_fa2_C and libattn_kernels_xe_2.so build source
native base	4772f727590c51b72add79350b913d098cf67872	_C, _xpu_C, grouped GEMM, and their helper DSOs
xpumem source	18a44f440ca3ac2006d5ba19cd12ccca0a0c9982	xpumem_allocator build source

The supplemental attention bundle matters. The measured _vllm_fa2_C.abi3.so had an absolute RUNPATH into its old build tree. Earlier repro code hashed a sibling attention DSO without proving that it was the file actually loaded. The hardened launcher puts the pinned kernel package first in LD_LIBRARY_PATH, imports all extensions, and verifies /proc/self/maps.

Restore and verify model payloads

The tracked manifest contains only the 32 release payload files. Hugging Face cache locks, metadata, and incomplete downloads are intentionally excluded.

repro/laguna-s-2.1-int4-b70-102tps-20260726/restore-models.sh \
  --download /mnt/fast-ai/llm-models/laguna-s-2.1

The downloader reads the token from /home/steve/.config/huggingface/token, requests these immutable revisions, and then hashes the complete payload:

• poolside/Laguna-S-2.1-INT4 at 4bbfc285f2f8b3b6b526274c133b7b17aae6c8cb;
• poolside/Laguna-S-2.1-DFlash-INT4 at 5e07c246915c86dc6920fead03d019989224f2ba.

For an existing copy, use restore-models.sh --verify.

Observed toolchain

• Ubuntu 24.04.4 LTS, kernel 7.0.0-28-generic, x86-64;
• Python 3.12.13;
• Intel oneAPI DPC++/C++ 2025.3.3;
• CC=/opt/intel/oneapi/compiler/2025.3/bin/icx;
• CXX=/opt/intel/oneapi/compiler/2025.3/bin/icpx;
• CMake Release;
• PyTorch 2.12.0+xpu, Triton XPU 3.7.1, oneCCL 2021.17.2;
• default paged-decode and chunk-prefill configs for the attention build.

The complete observed package listing is manifests/pip-freeze-observed.txt. Its editable vLLM and kernel metadata is stale and path-specific, so it is evidence, not an installable lock file. manifests/runtime-lock.json, Git HEADs, module origins, hashes, and loaded DSO maps are authoritative.

Source-equivalent native rebuild

Start with a Python 3.12 environment containing the exact package versions in manifests/runtime-lock.json, source the versioned oneAPI 2025.3 environment, and build with isolation disabled. The package’s standard full-build entry point is:

source /opt/intel/oneapi/compiler/2025.3/env/vars.sh
export CC=/opt/intel/oneapi/compiler/2025.3/bin/icx
export CXX=/opt/intel/oneapi/compiler/2025.3/bin/icpx
export CMAKE_BUILD_TYPE=Release
export MAX_JOBS=1
python setup.py build_ext --inplace

For the attention-runtime tree, also set:

export VLLM_PAGED_DECODE_CONFIG=paged_decode_default.conf
export VLLM_CHUNK_PREFILL_CONFIG=chunk_prefill_default.conf

Build each module from the source commit named in the provenance table, then place the resulting modules and helper DSOs together under one vllm_xpu_kernels/ directory. Do not reuse the historical absolute RUNPATH as a dependency mechanism. Put that directory first in LD_LIBRARY_PATH.

Because this is reconstructed rather than the missing original build invocation, its hashes are expected to differ. Record a new runtime lock and run the full semantic gate. Never substitute new hashes into the sealed artifact-exact lane and call it the same environment.

This site is open source. Improve this page.