Comfy-Org/ComfyUI · error · ValueError

MiniMax Music3 prompt has {prompt_tokens} tokens; maximum is

Error message

MiniMax Music3 prompt has {prompt_tokens} tokens; maximum is {MAX_PROMPT_TOKENS}

What it means

Raised by MiniMax Music3's generate() when the tokenized prompt (input_ids sequence length) exceeds MAX_PROMPT_TOKENS (5000). The autoregressive LM has a hard context limit, so an over-long prompt is rejected up front rather than failing mid-generation. The count includes caption tokens, lyrics tokens, and special markup tokens.

Source

Thrown at comfy/ldm/minimax_music/ar.py:237

    def _sample_c0(self, hidden, cfg_scale, top_k, generator, vocab_mask):
        if self.model.pruned_lm_head:
            guided = self._guided_c0(self.model.lm_head_pruned(hidden).float(), cfg_scale, top_k)
            code = sample_topk(guided, top_k, generator)
            stop_token = 0
            offset = 1
        else:
            logits = self.model.lm_head(hidden).float()
            stop_token = SPECIAL_TOKEN_IDS["<|audio_end|>"]
            logits = logits.masked_fill(vocab_mask, -float("inf"))
            guided = self._guided_c0(logits, cfg_scale, top_k).masked_fill(vocab_mask, -float("inf"))
            code = sample_topk(guided, top_k, generator)
            offset = AUDIO_CODE_OFFSET
        return torch.where(code == stop_token, 0, code - offset), code, stop_token

    def generate(self, input_ids, seed, max_audio_frames, device, cfg_scale=CFG_SCALE, top_k=CFG_TOP_K):
        prompt_tokens = int(input_ids.shape[1])
        if prompt_tokens > MAX_PROMPT_TOKENS:
            raise ValueError(f"MiniMax Music3 prompt has {prompt_tokens} tokens; maximum is {MAX_PROMPT_TOKENS}")

        input_ids = input_ids.to(device)
        if comfy.model_management.should_use_bf16(device):
            execution_dtype = torch.bfloat16
        else:
            execution_dtype = torch.float32
        unconditioned = input_ids.clone()
        unconditioned[:, 1:-2] = SPECIAL_TOKEN_IDS["<|audio_cfg|>"]
        text_ids = torch.cat((input_ids, unconditioned), dim=0)
        if self.model.pruned_embedding:
            text_embeds = self.model.embed_tokens_prefill(text_ids, out_dtype=execution_dtype)
        else:
            text_embeds = self.model.embed_tokens(text_ids, out_dtype=execution_dtype)
        decode_limit = min(int(max_audio_frames), MAX_AUDIO_FRAMES)
        past = self.model.init_kv_cache(2, prompt_tokens + decode_limit + 1, device, execution_dtype)
        output = self.model(None, embeds=text_embeds, past_key_values=past, dtype=execution_dtype)
        last_hidden = output[0][:, -1]
        past = output[2]

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Shorten the lyrics or caption so the full tokenized prompt is at most 5000 tokens
2. Check prompt construction for accidental duplication of the caption/lyrics template
3. If you build prompts programmatically, count tokens with the same tokenizer and truncate lyrics first (they dominate length)

Example fix

# before
input_ids = tokenizer(build_prompt(caption, full_lyrics), add_special_tokens=False)
out = generate(input_ids, ...)
# after
input_ids = tokenizer(build_prompt(caption, full_lyrics), add_special_tokens=False)
assert input_ids.shape[1] <= MAX_PROMPT_TOKENS, input_ids.shape[1]
out = generate(input_ids[:, -MAX_PROMPT_TOKENS:], ...)

from comfy.ldm.minimax_music.ar import MAX_PROMPT_TOKENS
prompt_tokens = input_ids.shape[1]
if prompt_tokens > MAX_PROMPT_TOKENS:
    input_ids = input_ids[:, -MAX_PROMPT_TOKENS:]  # or truncate lyrics and re-tokenize

When it happens

Trigger: Calling generate(input_ids, ...) where input_ids.shape[1] > 5000 — typically a long caption plus extensive normalized lyrics.

Common situations: Very long lyrics (multiple verses in the prompt template), verbose captions, or a prompt-building bug that concatenates the prompt template multiple times.

Related errors

• MiniMax Music3 tokenizer mismatch for {token}: expected {exp
• MiniMax Music3 text encoder checkpoint is missing tokenizer_
• MiniMax Music3 tokenizer mismatch for {token}
• MiniMax Music3 generated zero audio frames
• Unknown patch method: {self.patch_method}

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/cdba0f13ee43d5a0. Report an issue: GitHub.