Skip to content

LLM inference in C/C++

C++ ggml
4.1 KB
Raw Blame
from __future__ import annotations

from typing import TYPE_CHECKING

if TYPE_CHECKING:
    from torch import Tensor

from .base import ModelBase, TextModel, get_model_architecture, gguf

from .deepseek import DeepseekV2Model
from .llama import LlamaModel


@ModelBase.register(
    "Mistral3ForConditionalGeneration",
    "Mistral3Model",
    "Ministral3ForCausalLM",
    "Ministral3Model",
)
@ModelBase.example("mistralai/Mistral-Small-3.1-24B-Instruct-2503", "hf-tiny-v2/tiny-random-Ministral3ForCausalLM", "nvidia/Nemotron-3-Embed-1B-BF16", "nvidia/Nemotron-3.5-Embed-1B-BF16")
class Mistral3Model(TextModel):
    class Ministral3Model(LlamaModel):
        model_arch = gguf.MODEL_ARCH.MISTRAL3

        def set_gguf_parameters(self):
            super().set_gguf_parameters()
            rope_params = self.rope_parameters
            if self.hparams.get("model_type") == "ministral3":
                assert rope_params, "ministral3 must have 'rope_parameters' config"
                assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
                self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
                self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])

    class Mistral3EmbedNemotronModel(Ministral3Model):
        model_arch = gguf.MODEL_ARCH.MISTRAL3

        def set_vocab(self):
            super().set_vocab()
            self.gguf_writer.add_add_bos_token(False)
            self.gguf_writer.add_add_eos_token(False)

        def get_vocab_base_pre(self, tokenizer) -> str:
            if self.origin_hf_arch == "Mistral3Model":
                return "nemotron-3.5-embed"
            return super().get_vocab_base_pre(tokenizer)

        def set_gguf_parameters(self):
            super().set_gguf_parameters()
            self._try_set_pooling_type()

    class Mistral4Model(DeepseekV2Model):
        model_arch = gguf.MODEL_ARCH.MISTRAL4
        skip_mtp = False # model contains no MTP layers, so no need to skip
        merge_expert = False # experts are already stacked as 3D

        def modify_tensors(self, data_torch, name, bid):
            if name.endswith(".down_proj") or name.endswith(".gate_up_proj"):
                name = name + ".weight"
            yield from super().modify_tensors(data_torch, name, bid)

    model_arch = gguf.MODEL_ARCH.MISTRAL3 # unused
    impl: TextModel

    def __init__(self, dir_model, *args, **kwargs):
        hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
        hf_arch = get_model_architecture(hparams, self.model_type)
        if hf_arch in ("Ministral3Model", "Mistral3Model"):
            text_config = hparams.get("text_config", hparams)
            if text_config.get("is_causal") is not False:
                raise NotImplementedError(f"Expected a Nemotron Embed model for {hf_arch}, but is_causal is not False")
        super().__init__(dir_model, *args, **kwargs)
        if self.hparams.get("model_type") == "mistral4":
            self.impl = Mistral3Model.Mistral4Model(dir_model, *args, **kwargs)
        elif self.hf_arch in ("Ministral3Model", "Mistral3Model"):
            self.impl = Mistral3Model.Mistral3EmbedNemotronModel(dir_model, *args, **kwargs)
        else:
            self.impl = Mistral3Model.Ministral3Model(dir_model, *args, **kwargs)

    def set_vocab(self):
        self.impl.set_vocab()

    def set_gguf_parameters(self):
        self.impl.set_gguf_parameters()

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
        yield from self.impl.modify_tensors(data_torch, name, bid)

    def prepare_tensors(self):
        self.impl.prepare_tensors()

    def write_vocab(self):
        self.impl.write_vocab()

    def write(self):
        self.impl.write()


@ModelBase.register("VoxtralRealtimeForConditionalGeneration")
@ModelBase.example("mistralai/Voxtral-Mini-4B-Realtime-2602")
class VoxtralRealtimeModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.MISTRAL3

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        self.gguf_writer.add_realtime_delay_tokens(self.hparams["default_num_delay_tokens"])