#!/usr/bin/env python3
# /// script
# dependencies = ["torch", "transformers", "gguf", "huggingface_hub", "sentencepiece", "numpy", "tiktoken"]
# ///
"""Convert sakthai-coder-browser to GGUF (F16)."""
import os, sys, json, numpy as np
from pathlib import Path

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from gguf import GGUFWriter, GGMLQuantizationType, MODEL_ARCH, get_tensor_name_map

MODEL_ID = "Nanthasit/sakthai-coder-browser"
OUTPUT_DIR = Path("gguf-output")
OUTPUT_DIR.mkdir(exist_ok=True)
GGUF_PATH = OUTPUT_DIR / "sakthai-coder-browser-f16.gguf"

print("Loading model...", flush=True)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.float16, device_map=None, low_cpu_mem_usage=True)
print("Loading tokenizer...", flush=True)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

c = model.config
n_blocks = c.num_hidden_layers

print(f"Arch: Qwen2, layers={n_blocks}, heads={c.num_attention_heads}, "
      f"kv={c.num_key_value_heads}, embd={c.hidden_size}, ff={c.intermediate_size}", flush=True)

gguf_writer = GGUFWriter(str(GGUF_PATH), "qwen2")
gguf_writer.add_block_count(n_blocks)
gguf_writer.add_context_length(c.max_position_embeddings)
gguf_writer.add_embedding_length(c.hidden_size)
gguf_writer.add_feed_forward_length(c.intermediate_size)
gguf_writer.add_head_count(c.num_attention_heads)
gguf_writer.add_head_count_kv(c.num_key_value_heads)
gguf_writer.add_layer_norm_rms_eps(c.rms_norm_eps)
gguf_writer.add_rope_dimension_count(c.hidden_size // c.num_attention_heads)
gguf_writer.add_rope_freq_base(getattr(c, 'rope_theta', 10000.0))

tmap = get_tensor_name_map(MODEL_ARCH.QWEN2, n_blocks)
# Remove .weight/.bias suffixes to match the GGUF mapping (keys are base names)
def lookup_gguf_name(hf_name: str) -> str | None:
    # Try exact name first
    if hf_name in tmap.mapping:
        return tmap.mapping[hf_name][1]
    # Try without .weight or .bias suffix, then re-append .bias for bias tensors
    is_bias = hf_name.endswith('.bias')
    base = hf_name.removesuffix('.weight').removesuffix('.bias')
    if base in tmap.mapping:
        gguf_name = tmap.mapping[base][1]
        return gguf_name + '.bias' if is_bias else gguf_name
    return None

state = model.state_dict()
matched = 0
for hf_name, tensor in state.items():
    gguf_name = lookup_gguf_name(hf_name)
    if gguf_name is None:
        print(f"  SKIP: no mapping for {hf_name}", flush=True)
        continue
    data = tensor.to(torch.float32).numpy()
    gguf_writer.add_tensor(gguf_name, data)
    matched += 1

print(f"Matched {matched}/{len(state)} tensors", flush=True)

# Tokenizer
gguf_writer.add_tokenizer_model("gpt2")
gguf_writer.add_tokenizer_pre("default")
vocab = tokenizer.get_vocab()
sorted_tokens = sorted(vocab.items(), key=lambda x: x[1])
tokens = [t for t, _ in sorted_tokens]
scores = [0.0] * len(tokens)
gguf_writer.add_token_list(tokens)
gguf_writer.add_token_scores(scores)
gguf_writer.add_bos_token_id(tokenizer.bos_token_id or 151643)
gguf_writer.add_eos_token_id(tokenizer.eos_token_id or 151643)
gguf_writer.add_unk_token_id(tokenizer.unk_token_id or 0)

print("Writing GGUF...", flush=True)
gguf_writer.write_header_to_file()
gguf_writer.write_kv_data_to_file()
gguf_writer.write_tensors_to_file()
gguf_writer.close()

size_gb = GGUF_PATH.stat().st_size / 1e9
print(f"Done: {GGUF_PATH} ({size_gb:.2f} GB)", flush=True)