---
title: '56228'
pageLayout: default
slug: thoughts/inference/vllm-prs/56228
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/56228.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#56228](https://github.com/aarnphm/aarnphm.github.io/issues/56228) \[Model\] DeepSeek-V4.1-Flash Model Definitions

[vllm-project/vllm#56228](https://github.com/vllm-project/vllm/pull/56228) · ywang96 · MERGED · created 2026-09-10T07:47:48Z · merged 2026-09-10T12:16:26Z

Split out from [#56214](https://github.com/aarnphm/aarnphm.github.io/issues/56214) — this PR includes **only** the changes under `vllm/models/`:

- Adds the new `vllm/models/deepseek_v4_1/` model definition package (nvidia, amd, and common code paths)
- Includes the associated modifications to `vllm/models/deepseek_v4/`

All other changes from [#56214](https://github.com/aarnphm/aarnphm.github.io/issues/56214) (kernels, rust parser, config, tests, etc.) are intentionally excluded.

**Non-duplication:** This is not a competing implementation — it is a scoped subset of [#56214](https://github.com/aarnphm/aarnphm.github.io/issues/56214) (same author group), split out so the model definitions can be reviewed independently.

**Rebase note:** `vllm/models/deepseek_v4/common/ops/fused_inv_rope_fp8_quant.py` conflicted with [#50176](https://github.com/aarnphm/aarnphm.github.io/issues/50176) (which migrated the kernel to the `VllmTritonJitKernel` warmup class) and [#56035](https://github.com/aarnphm/aarnphm.github.io/issues/56035) (PDL warmup fix). The resolution ports this change’s `QUANTIZE` toggle, `NOPE_DIM`-based rope offset, and vectorized UE8M0 scale packing onto the new class structure, and keeps the `launch_pdl=current_platform.is_arch_support_pdl()` fix.

**AI assistance:** This PR was prepared with AI assistance (branch split and conflict resolution).

**Testing:** No tests were run locally for this split (no CUDA/ROCm environment on the authoring machine). The branch content is byte-identical to [#56214](https://github.com/aarnphm/aarnphm.github.io/issues/56214) for all files except the rebased `fused_inv_rope_fp8_quant.py`; CI on this PR plus the kernel tests in [#56214](https://github.com/aarnphm/aarnphm.github.io/issues/56214) (`tests/kernels/test_fused_inv_rope_fp8_quant.py`) should be the source of truth.

---

## comments

**mergify** (2026-09-10T08:25:07Z):

Hi @ywang96, the pre-commit checks have failed. Please run:



```bash
uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-files
```

Then, commit the changes and push to your branch.

For future commits, `pre-commit` will run automatically on changed files before each commit.

**ywang96** (2026-09-10T08:57:03Z):

/ci run

**github-actions** (2026-09-10T08:57:52Z):

✅ Triggered [Buildkite CI #88092](https://buildkite.com/vllm/ci/builds/88092) for commit `68cd5e4410af`.

**ZJY0516** (2026-09-10T09:55:43Z):

/ci run

**github-actions** (2026-09-10T09:56:01Z):

✅ Triggered [Buildkite CI #88099](https://buildkite.com/vllm/ci/builds/88099) for commit `b5e138f8c012`.

