vllm-canon 0.1.0

vLLM plugin: out-of-tree registration of canon-layer architectures (e.g. LlamaCanonForCausalLM from PhysicsLM4)

vllm-canon

An out-of-tree vLLM plugin that adds support for the LlamaCanonForCausalLM architecture — the "canon layer" variant of Llama introduced in Zeyuan Allen-Zhu's PhysicsLM4 / Canon Layers work.

A canon layer is a depthwise causal short convolution (kernel=4 by default) inserted at up to four positions in each decoder block:

• canonA — on the residual stream after input_layernorm, before attention
• canonB — on the fused qkv stream before RoPE
• canonC — on the residual stream after post_attention_layernorm, before MLP
• canonD — on the fused gate_up stream before silu * mul

Install

pip install vllm-canon

After install, vLLM auto-discovers the plugin via its vllm.general_plugins entry point.

Use

Pass trust_remote_code=True so HuggingFace autoloads the custom LlamaCanonConfig from your model directory:

from vllm import LLM, SamplingParams

llm = LLM(
    model="/path/to/your/canon-model",
    trust_remote_code=True,
    tensor_parallel_size=1,
    dtype="bfloat16",
    enforce_eager=True,
)
print(llm.generate(["hello"], SamplingParams(temperature=0, max_tokens=32))[0].outputs[0].text)

Or start a vLLM server:

vllm serve /path/to/your/canon-model \
  --trust-remote-code --tensor-parallel-size 1 --dtype bfloat16 \
  --enforce-eager --port 8000 --served-model-name canon

What the plugin does

• Registers LlamaCanonForCausalLM in ModelRegistry via the vllm.general_plugins entry point — no edits to the vLLM source tree.
• Rebuilds the Llama block with vLLM primitives (QKVParallelLinear, MergedColumnParallelLinear, paged attention, partial RoPE via partial_rotary_factor) and inserts the four canon convolutions at the HF reference positions.
• Each canon conv is a MambaBase with mamba_type="short_conv" so vLLM's V1 engine allocates a per-request (kernel-1, dim) rolling state alongside the KV cache. The model declares HasInnerState and IsHybrid so the engine plumbs that state correctly.
• The conv forward is written in pure PyTorch (F.conv1d for prefill, shift-append + dot for decode). The triton kernel (causal_conv1d_fn / causal_conv1d_update) produced state-update results that diverged from the reference in this setting; the canon width is tiny so the pure-torch path is fine.
• The HF checkpoint loads via vLLM's standard stacked-weight mapping: q_proj/k_proj/v_proj → qkv_proj, gate_proj/up_proj → gate_up_proj, canon weights by name. lm_head is tied to embed_tokens when tie_word_embeddings=True.

Limitations

• tensor_parallel_size=1 only. Canon B and canon D operate on fused QKV / gate_up streams; per-shard weight layouts under TP>1 need separate work.
• rope_version='huggingface' only. Lingua-style interleaved RoPE is not supported.
• enforce_eager=True recommended. The model class is not decorated with @support_torch_compile; adding it would require an explicit dynamic_arg_dims.

Compatibility

• vLLM >=0.15,<0.17 (tested on 0.15.1)
• transformers >= 4.57
• PyTorch >= 2.5
• Python >= 3.10

Parity

Verified against HuggingFace .generate() on the qwen1.5-0.5b-newtok-canon PhysicsLM4 checkpoint: 16/16 greedy tokens match for both a 1-token prompt (exercises the decode-path conv state update) and a 12-token prompt (exercises the prefill conv).

License

Apache 2.0. See LICENSE.