Public HF artifact now matches the validated clean baseline.
Bug 9b fix is included; prefill_chunk_size=16 workaround is no longer needed.
Debug instrumentation and experiment branches are removed from the active path.
Public HF smoke test passes from the HF URL, including the canonical France/Paris case.

Status

Text path: validated
Quantization: q4f16_1
Runtime target: webgpu
Model type: gemma4
Conversation template: gemma_instruction
Prefill chunk size: 1024
Debug instrumentation: removed
Known chunk16 workaround: not required

Files

libs/gemma-4-E2B-it-q4f16_1-MLC-webgpu.wasm: validated WebGPU model library
mlc-chat-config.json: MLC runtime configuration
params_shard_*.bin: quantized parameter shards
tensor-cache.json: tensor metadata cache
tokenizer.json, tokenizer_config.json: tokenizer assets
release-manifest.json: file inventory with SHA-256 hashes

Usage

Chat

mlc_llm chat HF://welcoma/gemma-4-E2B-it-q4f16_1-MLC

WebLLM Integration

import { CreateMLCEngine } from "@mlc-ai/web-llm";

const repo = "https://huggingface.co/welcoma/gemma-4-E2B-it-q4f16_1-MLC";

const appConfig = {
  model_list: [
    {
      model: repo,
      model_id: "gemma-4-E2B-it-q4f16_1-MLC",
      model_lib: `${repo}/resolve/main/libs/gemma-4-E2B-it-q4f16_1-MLC-webgpu.wasm`,
      required_features: ["shader-f16"],
    },
  ],
};

const engine = await CreateMLCEngine("gemma-4-E2B-it-q4f16_1-MLC", {
  appConfig,
});

Notes

This is a custom MLC/WebLLM artifact, not an official mlc-ai release. The validated scope is Gemma 4 E2B text generation on WebGPU.