# Manifest

Everything needed to say what produced these numbers.

## Node

| | |
| --- | --- |
| host | `alexsu-dev-b200-0`, 8x NVIDIA B200 (180 GiB each), exclusive |
| driver | 580.159.03 |
| CUDA | 13.0 (nvcc V13.0.88) |
| date | 2026-09-23 UTC |

## Runtime stack

One environment for everything: the conda env `dev`, Python 3.12.12.

| package | version |
| --- | --- |
| torch | 2.11.0+cu130 |
| transformers | 5.14.1 |
| flash-attn | 2.8.3 |
| flashinfer | 0.6.14 |
| triton | 3.6.0 |

Unlike a version-ablation study there is no second leg here, so this stack is a
description rather than an independent variable: both the baseline and the
candidate run inside it, and every speedup is a within-stack ratio.

## Code

| repo | commit | local changes |
| --- | --- | --- |
| [`Snowflake-AI-Research/fastkernels`](https://github.com/Snowflake-AI-Research/fastkernels) | `9acebaa` | **none** -- verified clean checkout |
| [`sfc-gh-goliaro/vanilla-agents-fastkernels`](https://github.com/sfc-gh-goliaro/vanilla-agents-fastkernels) | `0a05814` | one, `scripts/run.sh`, shipped as `patches/run.sh.patch` |

The FastKernels checkout being unmodified matters: the baselines the agent is
scored against, the tolerance checker, and the bench harness are all stock. It
was cloned fresh for this experiment specifically to keep it that way, separate
from the working copy used for an unrelated vLLM study on the same node.

## Agent

| | |
| --- | --- |
| CLI | `codex-cli` 0.156.1, prebuilt `x86_64-unknown-linux-musl` binary from the GitHub release |
| model | `openai-gpt-5.6-sol` |
| reasoning effort | high |
| API | Snowflake Cortex (Snowhouse), `responses` wire API |
| invocation | `codex exec --yolo --skip-git-repo-check -m openai-gpt-5.6-sol` |
| prompt | stock `prompts/optimize.md` from the harness, unmodified |

**The model id deviates from the harness default and this matters for
comparability.** `run.sh` defaults to `-m openai-gpt-5.6-sol:high`. The Cortex
endpoint rejects that with `unknown model`; only the bare `openai-gpt-5.6-sol`
resolves. The `:high` suffix was therefore moved out of the model id and into
`model_reasoning_effort = "high"` in `patches/codex-config.toml`, which the CLI
reports back as `reasoning effort: high` at session start. The intent is the
same; the mechanism is not the harness default.

Three further pieces of codex configuration were needed to reach Cortex at all,
all in `patches/codex-config.toml`:

- a custom `model_provider`, because codex 0.156.1 ignores `OPENAI_BASE_URL`,
  which is the only thing `run.sh` sets. Without this every request goes to
  `api.openai.com` and 401s.
- `model_catalog_json`, pointing at a catalog entry for `openai-gpt-5.6-sol`.
  Absent metadata makes codex fall back to defaults that enable the web-search
  tool, which Cortex rejects outright.
- `web_search = "disabled"` -- a top-level enum, not the `[tools] web_search`
  boolean, which is accepted by the config parser but has no effect here.

## Harness settings

Stock defaults except where noted.

| setting | value | source |
| --- | --- | --- |
| mode | `all` (independent, then sequential, then compare) | default |
| targets | every L1-L3 operator in the default capture: **133** | default |
| levels | 1,2,3 | default |
| concurrent agent sessions | 12 | default `MAX_AGENT` |
| session timeout | 5 h | default `AGENT_TIMEOUT` |
| GPU lease timeout | 15 min | default `BENCH_TIMEOUT` |
| winner gate | every non-skipped scenario `PASSED` **and** geomean >= 1.0 | default |
| `FK_DIR` | the clean fastkernels checkout | set explicitly |
| `AGENT_ARGS` | overridden, see above | **not** default |

Agents do not hold a GPU while thinking; they lease one through `with_gpu.py`
only to run `validate.py`, `ncu` or `nsys`.

## Captured inputs

Shapes and dtypes come from `captures/default/b200` in
`sfc-gh-goliaro/fastkernels-results`, downloaded by the harness on first run.
Its own provenance is `capture_20260821-081742`. These are recorded from real
model executions, not synthesised, which is the point of the capture mechanism
for data-dependent operators such as MoE routing.

## Credentials

The Cortex token has an 18 h lifetime and was rotated once mid-run, at
2026-09-23 18:24 UTC, between the sequential agent phase and the final scoring.
`run.sh` re-reads `$HOME/.snowhouse-pat` before every agent session, so no
restart was needed and no session ran with an expired token. No 401s appear in
any driver log.

## Archive contents

| path | what it is |
| --- | --- |
| `REPORT.md` | the analysis; `README.md` is generated from it |
| `CANONICAL.md` | which of the three scoring executions is authoritative, and where the artifacts are not self-consistent |
| `MANIFEST.md` | this file |
| `SOURCE_NOTES.md` | working documents that were deliberately not shipped, and why |
| `results/ind/` | standalone bench over all 133 merged candidates, the winner selection, per-target agent metrics |
| `results/comp/` | composed bench over the independent winners |
| `results/seq/` | composed bench over the sequential winners, plus the per-level mix benches and gating decisions |
| `results/compare.json` | the harness's own independent-vs-sequential comparison |
| `control/` | `fastkernels eval --self-test` on Llama-3.1-8B: the baseline-vs-baseline control that establishes the noise floor |
| `patches/` | the one harness change, the codex configuration, the scenario tables and driver scripts written for this run |
| `logs/` | the three driver logs |
| `logs-agents.tar.gz` | 219 per-target codex session logs, 107 MB uncompressed |
| `make_tables.py` | regenerates every table in `REPORT.md` from `results/` |
| `make_card.py` | regenerates `README.md` from `REPORT.md` |

The per-target agent workspaces (7,628 intermediate kernel snapshots, 5.8 GB)
are **not** shipped here; see `SOURCE_NOTES.md`.

## Local changes: `scripts/run.sh`

Two fixes, both required to get a complete run, both in `patches/run.sh.patch`.
Neither touches measurement -- they only stop the driver from aborting.

**1. `fastkernels bench` exiting non-zero aborts the driver.** `run_bench()`
called it bare under `set -euo pipefail`. Any failing candidate scenario makes
the command exit 1, which killed the driver one line before winner selection.
At 133 targets a clean sweep is not realistic, and the harness's own README
calls the standalone bench "diagnostic only" -- the gate is applied afterwards
by `select_fk_winners.py` reading the JSON. The fix distinguishes "some
scenarios failed" from "the command crashed": if the output JSON exists, carry
on. The same treatment is applied to `run_e2e()`.

**2. `find ... | grep -q .` misreports a large winner set as empty.** Both
`run.sh:559` and `:569` guard on
`! find "$dir" -name '*.py' | grep -q .`. `grep -q` exits at the first match and
closes the pipe, `find` takes SIGPIPE, and under `set -o pipefail` the pipeline
reports failure -- so 119 winners were read as zero and the composed bench and
e2e were silently skipped. It does not trigger with a handful of winners, which
is why the 5-target smoke run did not catch it. Replaced with
`[[ -z "$(find ... -print -quit)" ]]`, which stops `find` itself.

Both are worth upstreaming; the driver logs in `logs/` are the evidence.