---
license: other
license_name: per-artifact
license_link: https://huggingface.co/Pegainfer/kern-qwen38-sm103/blob/main/README.md
tags:
  - kern
  - cuda
  - cubin
  - sm103
  - qwen3.8
---
# kern-qwen38-sm103

Kernel artifacts and manifests for running **Qwen3.8-27B** (plain decode and
its DFlash2 speculative draft) on one **GB300 (sm_103a)** with
[kern](https://github.com/pegainfer-project/kern), a model-agnostic GPU
runtime. The runtime contains no model; everything it needs to know about
this one is in the manifest.

## Layout

- `cubins/<name>-<sha12>.cubin` — every device module the manifests pin, one
  flat content-addressed directory. The runtime resolves modules by sha256
  only; the name is a label and the sha prefix keeps every version of a
  kernel side by side. New builds are added, nothing is renamed or removed.
- `manifests/qwen3.8-27b.json` — the decode manifest (`prefill`, `decode`,
  `decode_batch` programs; full attention on TRTLLM-GEN, GDN layers on
  vLLM's Triton kernels plus kern's fused decode kernels). `decode` and
  `decode_batch` declare `graph` (driven through a captured CUDA graph);
  `prefill` runs launch by launch, up to 8192 tokens a chunk (`tokens`).
- `manifests/qwen3.8-27b-dflash2.json` — the same model with its DFlash2
  draft: draft, verify and round programs, zero extra runtime code.

Both manifests match `examples/` in the kern repository at the commit named
in each upload's commit message. They are manifest schema 5 and need kern
v0.2.0 or newer (`curl -fsSL https://kern-baa.pages.dev/install.sh | sh`
installs `kern` and `kern-serve` together); v0.1.0 refuses the file.

## Not hosted here

- **Weights and tokenizer.** They are the models' own Hugging Face
  checkpoints, unmodified: [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B)
  and, for the draft, [incoai/Qwen3.8-27B-DFlash2](https://huggingface.co/incoai/Qwen3.8-27B-DFlash2).
  Each weight buffer in the manifest names the checkpoint tensors (or the
  rectangles of them) that fill it, so the runtime assembles qkv / gate_up
  and the draft's fused buffers straight from the safetensors shards, and a
  `load` program computes the rest (rope tables, Gemma norm offsets) on the
  device.

## Run

```bash
hf download Pegainfer/kern-qwen38-sm103 --local-dir kern-qwen38
hf download Qwen/Qwen3.8-27B --local-dir qwen3.8-27b                  # the weights, as published
kern run --manifest kern-qwen38/manifests/qwen3.8-27b.json \
         --kernels  kern-qwen38/cubins \
         --weights  qwen3.8-27b \
         --gpu 0 --prompt "The capital of France is" --steps 32
```

For the speculative manifest add the draft checkpoint as a second
`--weights` (`hf download incoai/Qwen3.8-27B-DFlash2 --local-dir dflash2`,
then `--weights qwen3.8-27b --weights dflash2`); its 8-row draft / verify
round is the default step.

The same three flags serve it over an OpenAI-compatible endpoint:

```bash
kern-serve --manifest kern-qwen38/manifests/qwen3.8-27b.json --kernels kern-qwen38/cubins \
           --weights qwen3.8-27b --served-model-name qwen3.8-27b
```

`kern test --reference <other manifest>` attests one manifest against
another on logits; `kern bench` measures programs in isolation.

## Provenance and licenses

| artifacts | origin | license |
|---|---|---|
| `argmax`, `attn_prep`, `copy_rows`, `dflash_conv`, `dflash_select`, `embedding`, `gdn_advance`, `gdn_decode`, `gemm16`, `gemma_rms_norm`, `line_copy`, `sigmoid_mul`, `silu_mul`, `spec_round`, `topk_row`, `weight_prep` | kern's handwritten kernels, `tools/kernels-src/*.cu` in the kern repo, `nvcc -cubin -arch=sm_103a` | same as kern |
| `chunk_h`, `chunk_o`, `cumsum`, `kkt`, `recompute`, `solve_tril`, `post_conv`, `conv_fwd`, `conv_update_spec`, `recurrent_spec`, `layer_norm`, `mrope`, `cache`, `reduce`, `unified`, `unified_noncausal`, `d_cache`, `d_rms_norm`, `d_rope`, `vllm_activation` | captured from vLLM 0.28's Triton JIT and CUDA extension on GB300 (the GDN ops are the flash-linear-attention kernels vLLM vendors) | Apache-2.0 (vLLM); the vendored FLA ops are MIT |
| `fmhaSm100fKernel_…SwapsAbForGen` (×2), `fmhaSm103aKernel_…PersistentContext` | NVIDIA TensorRT-LLM's generated attention kernels (trtllm-gen), copied byte for byte from NVIDIA's public artifactory pin `158f6fa11ef139a098cfddcdddce73ca99d164ad/fmha/trtllm-gen/` (the same files FlashInfer downloads and ships in `flashinfer-cubin`, and TensorRT-LLM checks into its tree); sha256 pinned in the kern repo's `tools/trtllm-gen/artifacts.json` | Apache-2.0: the `LICENSE` NVIDIA publishes next to the cubins, copied to `licenses/LICENSE.trtllm-gen.txt` |

Toolchain of the handwritten kernels the base manifest pins (`attn_prep`,
`gdn_decode`, `gemm16`, `gemma_rms_norm-92dda…`, `sigmoid_mul-2b07…`,
`silu_mul`, `weight_prep`): CUDA 13.0 nvcc (`cuda_13.0.r13.0/compiler.36424714_0`). A
different nvcc yields a different cubin and therefore a different sha; the
manifest pins the build that was present when it was generated, and whether
two builds agree numerically is `kern test`'s verdict, not the name's.
The older handwritten builds were made before the toolchain was recorded.