maya1 GGUF example. To use this, first launch the maya1 gguf using llama-cpp or LM Studio. Then the example below should 'just work'.

example.py

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "httpx>=0.28.1",
#     "snac>=1.2.1",
#     "torch>=2.10.0",
#     "transformers>=5.2.0",
# ]
# ///
import httpx
import torch
from snac import SNAC
from transformers import AutoTokenizer


class Maya1Helpers:
    CODE_START_TOKEN_ID = 128257
    CODE_END_TOKEN_ID = 128258
    CODE_TOKEN_OFFSET = 128266
    SNAC_MIN_ID = 128266
    SNAC_MAX_ID = 156937
    SNAC_TOKENS_PER_FRAME = 7
    SNAC_OFFSET = 4096

    SOH_ID = 128259
    EOH_ID = 128260
    SOA_ID = 128261
    BOS_ID = 128000
    TEXT_EOT_ID = 128009

    def __init__(self):
        self.tokenizer = AutoTokenizer.from_pretrained(
            "maya-research/maya1", trust_remote_code=True
        )
        self.snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval()

    def build_prompt(self, text: str, description: str):
        """Build formatted prompt for Maya1."""
        soh_token = self.tokenizer.decode([self.SOH_ID])
        eoh_token = self.tokenizer.decode([self.EOH_ID])
        soa_token = self.tokenizer.decode([self.SOA_ID])
        sos_token = self.tokenizer.decode([self.CODE_START_TOKEN_ID])
        eot_token = self.tokenizer.decode([self.TEXT_EOT_ID])
        bos_token = self.tokenizer.bos_token

        formatted_text = f'<description="{description}"> {text}'

        prompt = (
            soh_token + bos_token + formatted_text + eot_token + eoh_token + soa_token + sos_token
        )

        return prompt

    def _extract_snac_codes(self, token_ids: list) -> list:
        """Extract SNAC codes from generated tokens."""
        try:
            eos_idx = token_ids.index(self.CODE_END_TOKEN_ID)
        except ValueError:
            eos_idx = len(token_ids)

        snac_codes = [
            token_id
            for token_id in token_ids[:eos_idx]
            if self.SNAC_MIN_ID <= token_id <= self.SNAC_MAX_ID
        ]

        return snac_codes

    def _unpack_snac(self, snac_tokens):
        # Decode SNAC tokens to audio frames
        frames = len(snac_tokens) // self.SNAC_TOKENS_PER_FRAME
        codes = [[], [], []]
        for i in range(frames):
            s = snac_tokens[i * self.SNAC_TOKENS_PER_FRAME : (i + 1) * self.SNAC_TOKENS_PER_FRAME]
            codes[0].append((s[0] - self.SNAC_MIN_ID) % self.SNAC_OFFSET)
            codes[1].extend(
                [
                    (s[1] - self.SNAC_MIN_ID) % self.SNAC_OFFSET,
                    (s[4] - self.SNAC_MIN_ID) % self.SNAC_OFFSET,
                ]
            )
            codes[2].extend(
                [
                    (s[2] - self.SNAC_MIN_ID) % self.SNAC_OFFSET,
                    (s[3] - self.SNAC_MIN_ID) % self.SNAC_OFFSET,
                    (s[5] - self.SNAC_MIN_ID) % self.SNAC_OFFSET,
                    (s[6] - self.SNAC_MIN_ID) % self.SNAC_OFFSET,
                ]
            )
        return codes

    def get_audio_from_prompt_output(self, output: str):
        levels = self._unpack_snac(self._extract_snac_codes(self.tokenizer.encode(output)))
        codes_tensor = [torch.tensor(level, dtype=torch.long).unsqueeze(0) for level in levels]
        with torch.inference_mode():
            z_q = self.snac_model.quantizer.from_codes(codes_tensor)
            audio = self.snac_model.decoder(z_q)[0, 0].cpu().numpy()
        return audio


API_URL = "http://127.0.0.1:1234/v1/completions"
HEADERS = {"Content-Type": "application/json"}
maya1 = Maya1Helpers()
prompt = maya1.build_prompt(text, description)
output = httpx.post(
    API_URL,
    headers=HEADERS,
    json={
        "model": "maya1",
        "prompt": prompt,
    },
    timeout=None,
)
output = output.json()["choices"][0]["text"]
audio = maya1.get_audio_from_prompt_output(output)