Skip to content

Update AGENT.md to match the current engine state#560

Open
Cioscos wants to merge 1 commit into
antirez:mainfrom
Cioscos:docs/update-agent-md
Open

Update AGENT.md to match the current engine state#560
Cioscos wants to merge 1 commit into
antirez:mainfrom
Cioscos:docs/update-agent-md

Conversation

@Cioscos

@Cioscos Cioscos commented Jul 13, 2026

Copy link
Copy Markdown

AGENT.md had drifted behind the code: it predates the ROCm/Strix Halo backend, the Pro model variant, MTP speculative decoding, and most of the current binaries and test tooling. This PR brings it back in sync while keeping it what it is — a short philosophy/rules document — without duplicating the build/test mechanics already covered by CLAUDE.md and CONTRIBUTING.md.

Changes

  • Intro: the engine is now described as DeepSeek V4 Flash/Pro specific, with the shape selected from the GGUF at ds4_engine_open() time (the code has had DS4_VARIANT_FLASH/DS4_VARIANT_PRO shape tables for a while).
  • Goals:
    • Added ROCm to the list of inference paths that must not be broken by fixes to other parts of the code.
    • Added a goal for MTP speculative decoding correctness: speculative acceptance must never change the tokens a non-speculative run would produce.
  • Safety: hardened the macOS note from "avoid large CPU inference runs" to a hard rule — never run the CPU inference path on macOS (kernel VM bug, can require a reboot); use make cpu there only to confirm the build compiles. This matches the current guidance in CLAUDE.md.
  • Layout:
    • Listed the five binaries the build produces (ds4, ds4-server, ds4-bench, ds4-eval, ds4-agent).
    • Added the source files/dirs that were missing: ds4_cuda.cu/ds4_rocm.cu (+ ds4_gpu.h, rocm/*.cuh), ds4_agent.c (+ ds4_web.c), ds4_distributed.c, ds4_ssd.c, ds4_kvstore.c (+ rax.c), ds4_bench.c/ds4_eval.c, gguf-tools/.
  • Testing:
    • Mentioned the focused ds4_test sub-checks (--server, --logprob-vectors, --long-context, --tool-call-quality, --metal-kernels), pointing at CONTRIBUTING.md for details.
    • Extended checklist item 4 from "CUDA" to "CUDA or ROCm", asking for access to the relevant machine.

Docs-only change; no code touched.

🤖 Generated with Claude Code

AGENT.md predated several major additions to the engine. Bring it back in
sync without duplicating the build/test mechanics already documented in
CLAUDE.md and CONTRIBUTING.md:

- Describe the engine as Flash/Pro specific (shape selected from the GGUF).
- Add ROCm to the list of inference paths that must keep working, and to
  the per-change testing checklist alongside CUDA.
- Add an MTP/speculative-decoding correctness goal.
- Harden the macOS safety note: never run the CPU inference path there.
- List the five built binaries and the missing source files/dirs in Layout
  (CUDA/ROCm backends, agent, distributed, SSD streaming, KV store, bench,
  eval, gguf-tools).
- Point Testing at the ds4_test sub-checks and CONTRIBUTING.md.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant