Multi Repo Dsv4flash Setup

Concrete guide to benchmark DeepSeek-V4 (dsv4) FLASH on NPU against the current worktree's simpler — the three decode attentions (swa/csa/hca) and the distributed MoE, all in pypto-lib's models/deepseek/v4/. Defers cross-repo clone/install to multi-repo-setup, then gives the exact PYPTO_BENCH run commands and the hard-won config that makes the numbers reproducible: pin ptoas 0.48 (0.50/ptoas-bin cannot compile csa), pin --start-pos 8192, edit moe.py for 16 experts/card, read fast_effective_us not effective_us, pin an even die, and apply the eager-init idempotency guard so distributed MoE runs

49 stars 102 forksGit & Code Review

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add hw-native-sys/simpler --skill multi-repo-dsv4flash-setup
Manual install
git clone --depth 1 https://github.com/hw-native-sys/simpler.git /tmp/simpler
mkdir -p ~/.claude/skills
cp -r /tmp/simpler/.claude/skills/multi-repo-dsv4flash-setup ~/.claude/skills/multi-repo-dsv4flash-setup

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# DeepSeek-V4 FLASH on NPU: attention + MoE effective-time benchmark

This skill is the **dsv4-flash-specific concrete guide**. It benchmarks the
four kernels that live in **pypto-lib** at `models/deepseek/v4/` against the
current worktree's simpler:

| kernel | file | measures | dies |
| ------ | ---- | -------- | ---- |
| SWA attention | `decode_attention_swa.py` | decode attn effective_us | 1 |
| CSA attention | `decode_attention_csa.py` | decode attn effective_us | 1 |
| HCA attention | `decode_attention_hca.py` | decode attn effective_us | 1 |
| MoE (distributed) | `moe.py` | decode MoE fast_effective_us | ≥2 (EP) |

It sits on top of [`multi-repo-setup`](../multi-repo-setup/SKILL.md), which
owns the generic repo-graph + clone + install steps this skill does not
repeat. Everything below is the dsv4-specific run recipe and the five
configuration facts that make the effective-time numbers **reproducible to
~1%** instead of wandering ±18% run-to-run.

## 1. Setup — defer to multi-repo-setup

Run [`multi-repo-setup`](../multi-repo-setup/SKILL.md) first — it clones
pto-isa / pypto / pypto-lib under `build/`, exports the toolchain env, and
installs the simpler you want (worktree or main). Then, from the worktree:

```bash
source .venv/bin/activate
eval "$(pypto-setup --export)"            # ASCEND_HOME_PATH, gcc, PATH
```

Verify the loaded simpler is the worktree's, not a user-site shadow:
`python -c "import simpler; print(simpler.__file__)"`.
Read the full skill on GitHub

Details

Skill name
multi-repo-dsv4flash-setup
Path
.claude/skills/multi-repo-dsv4flash-setup/SKILL.md
Repo updated
18 November 2026

FAQ

What does the Multi Repo Dsv4flash Setup skill do?

Concrete guide to benchmark DeepSeek-V4 (dsv4) FLASH on NPU against the current worktree's simpler — the three decode attentions (swa/csa/hca) and the distributed MoE, all in pypto-lib's models/deepseek/v4/. Defers cross-repo clone/install to multi-repo-setup, then gives the exact PYPTO_BENCH run commands and the hard-won config that makes the numbers reproducible: pin ptoas 0.48 (0.50/ptoas-bin cannot compile csa), pin --start-pos 8192, edit moe.py for 16 experts/card, read fast_effective_us not effective_us, pin an even die, and apply the eager-init idempotency guard so distributed MoE runs

How do I install the Multi Repo Dsv4flash Setup skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add hw-native-sys/simpler --skill multi-repo-dsv4flash-setup. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Multi Repo Dsv4flash Setup popular?

It lives in hw-native-sys/simpler, which has 49 GitHub stars and 102 forks.

Related skills

Benchmark

49

hw-native-sys/simpler

Benchmark runtime performance on hardware. If the current branch has commits ahead of upstream/main or uncommitted changes, compares against the fork point (merge-base). Otherwise benchmarks current state only. Use when the user asks to benchmark, measure performance, or compare latency.

Git & Code Review102

Checkout PR

49

hw-native-sys/simpler

Add a remote for a given PR's head repository and checkout a local pr-xxx-work branch. Use when the user wants to work on someone else's PR locally.

Git & Code Review102

Clean Branches

49

hw-native-sys/simpler

Remove stale local and remote git branches that have been merged into main. Detects regular merges and squash merges. Cleans up fork remote branches and prunes tracking refs. Use when the user asks to clean up branches, remove merged branches, or tidy up git.

Git & Code Review102

Multi Repo Qwen Setup

49

hw-native-sys/simpler

Concrete qwen3-14B guide on NPU against the current worktree's simpler. Points at simpler's own in-repo qwen3_14b_decode example (zero cross-repo), then covers the two cross-repo paths — the pypto-serving runner (prefill/decode TPOT) and the pypto-lib decode_fwd kernel (device decode TPOT + scheduler DFX). Defers cross-repo clone/install to multi-repo-setup, then gives the exact run commands (batch-1 and batch-16), the prefill ring/timeout gotchas that otherwise surface as 507018, how to read the per-token timing layers, and the kernel-level device-timing / swimlane flow. Invoke when running q

Git & Code Review102

Maintainer

49

hw-native-sys/simpler

Maintain GitHub Issues and PRs through an action table, focused user decisions and recurring monitoring or a requested single pass. Use for repository, Issue, track or personal scopes, optionally filtering PR target branches. Scope does not grant repository authority; continuous monitoring requires a working wake mechanism.

Git & Code Review102

Multi Repo Setup

49

hw-native-sys/simpler

Set up a cross-repo investigation when a workload from another repo (pypto, pypto-lib, pypto-serving, etc.) needs to be run, especially when you want to swap in simpler-main HEAD or the current worktree's simpler instead of the version that repo pins. Documents the repo graph (who imports who), clones-or-updates each external repo every invocation so stale local clones don't lie about CI parity, and points at each repo's own skills/docs for workload-specific steps. MUST invoke before chasing "X doesn't work on simpler" reports where X lives outside this repo.

Git & Code Review102

Browse the most popular Claude skills, all git & code review skills, or write your own with the skills generator.