b70-optimization-lab

Laguna S 2.1 published-convention 102.971 tok/s source snapshots

These files preserve the exact local vLLM and vLLM XPU-kernel source used by the approved four-B70 Laguna result (101.942 tok/s under conventional interval accounting). The original experiment branches point at upstream project remotes and are not assumed to be publicly pushable.

Component	Public upstream base	Record commit	Bundle SHA-256	Combined patch SHA-256
vLLM	`1d354c694eedf241b836319dba497345a5aa1167`	`e596ef1543466ae1a05e5bb8091f58872e2b18ba`	`56bb463925cf53ba20db81c980264fd2a993d1e76f15731ac5676674fa0b9126`	`23fc7766c1985c6237fcab12dae62096f759511dbc7c5dc7f9828fe93d4c6d6d`
vLLM XPU kernels	`11f42aa47ff51924b3d9527cfc2bfef5fd2d98e5`	`6f9dd3c3a7b1b677a992ca4f431a968408f9c816`	`7aeec0b55e7b40f26d993ddf6dfb978ec267da1ae1e89d9f54c6db4d80022ed3`	`0eb4337547160394ff4ebaa66782dd936fbad3fffeceb85af08832f6d99dbd69`
supplemental attention-runtime XPU source	`11f42aa47ff51924b3d9527cfc2bfef5fd2d98e5`	`906190641d708b8028018c5dde653e265c835348`	`8c53de764a0c903a5e627a1c7a54a9c3bb21b500479a644b86af4c83e948149e`	`0420ac1a6db050d7a695dcfa60e8dcec2286b72b1660509862587aa537a05e0d`

The .bundle files preserve the original commit graph after the listed public prerequisite. The .patch files are review-friendly binary-safe combined tree diffs from that base to the record commit. Both bundle prerequisites are contained in the corresponding upstream origin/main.

Restore vLLM from an upstream clone:

git fetch origin
git fetch /path/to/b70-optimization-lab/patches/laguna-s-2.1-xpu-b70/vllm-laguna-width12-dflash-fp8-102tps-record-20260726.bundle \
  experiment/laguna-width12-stack-clean-20260726:refs/heads/laguna-record
git checkout e596ef1543466ae1a05e5bb8091f58872e2b18ba

Restore the XPU kernels from an upstream clone:

git fetch origin
git fetch /path/to/b70-optimization-lab/patches/laguna-s-2.1-xpu-b70/vllm-xpu-kernels-laguna-width12-102tps-record-20260726.bundle \
  experiment/laguna-width12-router-clean-20260726:refs/heads/laguna-record
git checkout 6f9dd3c3a7b1b677a992ca4f431a968408f9c816

Restore the supplemental attention-runtime source:

git fetch /path/to/b70-optimization-lab/patches/laguna-s-2.1-xpu-b70/vllm-xpu-kernels-laguna-attention-runtime-906190641-20260726.bundle \
  experiment/laguna-s-2.1-fwht-20260721:refs/heads/laguna-attention-runtime
git checkout 906190641d708b8028018c5dde653e265c835348

The supplemental bundle closes an important provenance gap. The measured _vllm_fa2_C.abi3.so and libattn_kernels_xe_2.so were built from 906190641, while the final router/workspace source tree was 6f9dd3c3a. The measured _C, _xpu_C, grouped-GEMM/helper DSOs byte-match the earlier 4772f727590c51b72add79350b913d098cf67872 build, and xpumem_allocator came from 18a44f440ca3ac2006d5ba19cd12ccca0a0c9982. Both earlier commits are contained by the main kernel bundle. The standalone repro restores all of these worktrees, pins every native hash, and checks the actual loaded paths.

Validation performed before promotion:

both bundles pass git bundle verify in repositories containing their public prerequisites;
both combined patches pass git apply --check --reverse at their exact record commits;
bundle heads equal the record commits in the table;
all six file SHA-256 values were recomputed from the tracked snapshots.
the supplemental bundle and patch pass the same bundle-head, prerequisite, hash, and reverse-apply checks.

Result and reproduction entry points:

This site is open source. Improve this page.