| Parameter | Value |
|---|---|
| Source | utter-project/EuroLLM-1.7B-Instruct |
| Architecture | LlamaForCausalLM (llama) |
| Quantization | int4 palettization (group_size=32) |
| Context length | 512 |
| Compute units | CPU_AND_NE (ANE-safe, 0 GPU-pinned ops) |
| Converter | john-rocky/CoreML-LLM (MonolithicWrapper) |
| coremltools | 8.3 |
verify_ane.py result: 5152 ops inspected, 0 GPU-preferred ops — PASS1{
2 "architecture": "llama",
3 "hidden_size": 2048,
4 "num_hidden_layers": 24,
5 "num_attention_heads": 16,
6 "num_key_value_heads": 8,
7 "head_dim": 128,
8 "vocab_size": 128000,
9 "context_length": 512,
10 "rms_norm_eps": 1e-5,
11 "quantization": "int4",
12 "compute_units": "ALL"
13}<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant<s><|im_end|> (token id 4)LlamaTokenizer"llama" to select a buildPrompt function, verify it applies ChatML markers (<|im_start|> / <|im_end|>) rather than Llama-3 markers (<|begin_of_text|> / <|eot_id|>). The two are incompatible — using Llama-3 markers on EuroLLM will produce garbage output.model.mlpackage — monolithic CoreML model (embed + transformer + lm_head, stateful KV cache)model_config.json — metadata for Swift inference enginehf_model/ — downloaded HF snapshot (tokenizer files; model.safetensors excluded as redundant)