Skip to content

LLM inference in C/C++

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

import re
from typing import Any, Callable, Iterable, Sequence, TYPE_CHECKING

import numpy as np
import torch

if TYPE_CHECKING:
    from pathlib import Path
    from torch import Tensor

from .base import MmprojModel, ModelBase, ModelType, TextModel, get_model_architecture, gguf, logger

from .granite import GraniteHybridModel
from .qwen import DFlashModel


@ModelBase.register("LiLiCorrDraftModel")
class LiLiCorrDraftModel(DFlashModel):
    model_arch = gguf.MODEL_ARCH.DFLASH

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        prefix = "model.lilicorr."
        if not any(name.startswith(prefix) for name in self.model_tensors):
            raise ValueError("no lilicorr.* tensors found")
        dflash_config = self.hparams["dflash_config"]
        self.lilicorr_geometry = {
            "hidden_size": int(dflash_config["lilicorr_hidden_size"]),
            "candidate_topk": int(dflash_config["lilicorr_candidate_topk"]),
            "num_heads": int(dflash_config["lilicorr_num_heads"]),
            "num_layers": int(dflash_config["lilicorr_num_layers"]),
            "factor_dim": int(dflash_config["lilicorr_factor_dim"]),
            "mlp_hidden_size": self.model_tensors[prefix + "layers.0.mlp.up_proj.weight"]().shape[0],
        }
        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count + self.lilicorr_geometry["num_layers"] + 1)
        block = int(dflash_config.get("block_size", self.hparams.get("block_size", 16)))
        topk = self.lilicorr_geometry["candidate_topk"]
        heads = self.lilicorr_geometry["num_heads"]
        nodes = (block - 1) * topk
        relative = gguf.LazyBase.to_eager(self.model_tensors.pop(prefix + "relative_slot_bias")()).float().cpu().numpy()
        same = gguf.LazyBase.to_eager(self.model_tensors.pop(prefix + "same_slot_bias")()).float().cpu().numpy()
        bias = np.empty((heads, nodes, nodes), dtype=np.float32)
        for q in range(nodes):
            q_slot = q // topk
            for k in range(nodes):
                k_slot = k // topk
                bias[:, q, k] = relative[:, q_slot - k_slot + block - 1]
                if q_slot == k_slot:
                    bias[:, q, k] += same
        self.model_tensors[prefix + "attention_bias.bias"] = lambda bias=bias: torch.from_numpy(bias)
        for name in ("slot_embedding", "rank_embedding"):
            gen = self.model_tensors.pop(prefix + name)
            self.model_tensors[prefix + name + ".weight"] = lambda gen=gen: gen().reshape(-1, self.lilicorr_geometry["hidden_size"])
        for i in range(self.lilicorr_geometry["num_layers"]):
            layer_prefix = prefix + f"layers.{i}.attn.in_proj"
            for suffix in ("weight", "bias"):
                self.model_tensors[layer_prefix + "." + suffix] = self.model_tensors.pop(layer_prefix + "_" + suffix)
        dflash_config["selector_rank"] = 0
        dflash_config["selector_top_k"] = 0

    def set_vocab(self):
        if self.target_model_dir is None:
            raise ValueError("LiLiCorr draft requires --target-model-dir")
        from . import get_model_class
        target_text_config = ModelBase.load_hparams(self.target_model_dir, False)
        target_text_config = {**target_text_config, **target_text_config.get("text_config", {})}
        target_arch = get_model_architecture(target_text_config, ModelType.TEXT)
        target_cls = get_model_class(target_arch)
        original_dir = self.dir_model
        self.dir_model = self.target_model_dir
        try:
            if target_arch == "NemotronHForCausalLM":
                self.is_moe = "num_experts_per_tok" in target_text_config
            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
        finally:
            self.dir_model = original_dir
        self.gguf_writer.add_mask_token_id(self.hparams["dflash_config"]["mask_token_id"])

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        dflash_config = self.hparams["dflash_config"]
        self.gguf_writer.add_lilicorr_hidden_size(int(self.lilicorr_geometry["hidden_size"]))
        self.gguf_writer.add_lilicorr_candidate_topk(int(self.lilicorr_geometry["candidate_topk"]))
        self.gguf_writer.add_lilicorr_num_heads(int(self.lilicorr_geometry["num_heads"]))
        self.gguf_writer.add_lilicorr_num_layers(int(self.lilicorr_geometry["num_layers"]))
        self.gguf_writer.add_lilicorr_factor_dim(int(self.lilicorr_geometry["factor_dim"]))
        self.gguf_writer.add_lilicorr_mlp_hidden_size(int(self.lilicorr_geometry["mlp_hidden_size"]))
        self.gguf_writer.add_lilicorr_vector_eps(float(dflash_config["lilicorr_vector_eps"]))
        self.gguf_writer.add_lilicorr_logit_scale(float(dflash_config["lilicorr_logit_scale"]))

    def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
        if name.startswith("model.lilicorr.layers."):
            parts = name.split(".", 4)
            if not parts[4].startswith("attn.in_proj."):
                parts[3] = str(self.block_count + int(parts[3]))
            name = ".".join(parts)
        elif name.startswith("model.lilicorr.feature_norm."):
            feature_bid = self.block_count + self.lilicorr_geometry["num_layers"]
            name = name.replace("model.lilicorr.feature_norm.", f"model.lilicorr.feature_mlp.{feature_bid}.norm.")
        elif name.startswith("model.lilicorr.feature_mlp."):
            feature_bid = self.block_count + self.lilicorr_geometry["num_layers"]
            name = name.replace("model.lilicorr.feature_mlp.", f"model.lilicorr.feature_mlp.{feature_bid}.")
        return super().map_tensor_name(name, try_suffixes)

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if name.startswith("model.lilicorr."):
            yield self.map_tensor_name(name), data_torch
            return
        yield from super().modify_tensors(data_torch, name, bid)

    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
        if name.startswith("model.lilicorr."):
            return gguf.GGMLQuantizationType.F32
        return super().tensor_force_quant(name, new_name, bid, n_dims)


@ModelBase.register(
    "NemotronH_Nano_VL_V2",
    "RADIOModel",
)
@ModelBase.example("nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16")
class NemotronNanoV2VLModel(MmprojModel):
    # ViT-Huge architecture parameters for RADIO v2.5-h
    _vit_hidden_size = 1280
    _vit_intermediate_size = 5120
    _vit_num_layers = 32
    _vit_num_heads = 16

    def get_vision_config(self) -> dict[str, Any] | None:
        # RADIO config doesn't have standard ViT parameters, so they need to be constructed manually
        vision_config = self.global_config.get("vision_config")
        if vision_config is None:
            return None
        # Add ViT-H parameters
        vision_config = {
            **vision_config,
            "hidden_size": self._vit_hidden_size,
            "intermediate_size": self._vit_intermediate_size,
            "num_hidden_layers": self._vit_num_layers,
            "num_attention_heads": self._vit_num_heads,
            "image_size": self.global_config.get("force_image_size", 512),
        }
        return vision_config

    def get_audio_config(self) -> dict[str, Any] | None:
        return self.global_config.get("sound_config")

    def set_gguf_parameters(self):
        if "image_mean" not in self.preprocessor_config:
            self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406]
        if "image_std" not in self.preprocessor_config:
            self.preprocessor_config["image_std"] = [0.229, 0.224, 0.225]

        if self.hparams_audio is not None:
            self.has_vision_encoder = True
            self.has_audio_encoder = True
            self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["num_mel_bins"])
            self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
            self.gguf_writer.add_audio_subsampling_factor(self.hparams_audio["subsampling_factor"])
            self.gguf_writer.add_audio_conv_kernel_size(self.hparams_audio["conv_kernel_size"])
            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.PARAKEET)
            self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
        else:
            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)

        super().set_gguf_parameters()
        hparams = self.global_config
        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
        self.gguf_writer.add_vision_use_gelu(True)
        downsample_ratio = hparams.get("downsample_ratio", 0.5)
        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))

    def tensor_force_quant(self, name, new_name, bid, n_dims):
        if "sound_encoder" in name or new_name.startswith("mm.a."):
            if "bias" in new_name or "norm" in new_name:
                return gguf.GGMLQuantizationType.F32
            if "conv" in new_name and "weight" in new_name:
                return gguf.GGMLQuantizationType.F32

        return super().tensor_force_quant(name, new_name, bid, n_dims)

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        if (titem := super().filter_tensors(item)) is None:
            return None
        name, gen = titem

        if "input_conditioner" in name:
            return None

        # mtmd does not support video yet so skip tensors related to video.
        if "radio_model.model.patch_generator.video_embedder" in name:
            return None

        if not name.startswith(("vision_model.radio_model.model.", "mlp1.", "sound_encoder.", "sound_projection.")):
            return None

        if "patch_generator.pos_embed" in name:
            if not name.endswith(".weight"):
                name += ".weight"

        # num_batches is only used for training not inference.
        if "conv.norm" in name and "num_batches" in name:
            return None

        return name, gen

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
        if "patch_generator.pos_embed" in name:
            # Downsample position embeddings for fixed 512x512 image size
            import torch.nn.functional as F
            n_embd = self.hparams["hidden_size"]
            image_size = self.global_config.get("force_image_size", 512)
            patch_size = self.hparams["patch_size"]
            target_patches_per_side = image_size // patch_size  # 32
            max_patches_per_side = int((data_torch.shape[1]) ** 0.5)  # 128
            if target_patches_per_side != max_patches_per_side:
                # Reshape to grid, interpolate, flatten back
                data_torch = data_torch.reshape(1, max_patches_per_side, max_patches_per_side, n_embd)
                data_torch = data_torch.permute(0, 3, 1, 2).float()  # [1, n_embd, 128, 128]
                data_torch = F.interpolate(data_torch, size=(target_patches_per_side, target_patches_per_side),
                                           mode='bilinear', align_corners=True)
                data_torch = data_torch.permute(0, 2, 3, 1)  # [1, 32, 32, n_embd]
                data_torch = data_torch.reshape(1, target_patches_per_side * target_patches_per_side, n_embd)

        # Reshape linear patch embedding to conv2d format for ggml_conv_2d
        # From [n_embd, patch_size*patch_size*3] to [n_embd, 3, patch_size, patch_size]
        if "patch_generator.embedder" in name:
            patch_size = self.hparams["patch_size"]
            n_embd = self.hparams["hidden_size"]
            data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size)

        if "depthwise_conv.weight" in name:
            data_torch = data_torch.unsqueeze(-1)
            data_torch = data_torch.permute(3, 1, 0, 2).contiguous()

        if "pointwise_conv" in name and name.endswith(".weight"):
            if len(data_torch.shape) == 3 and data_torch.shape[2] == 1:
                data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])

        if "subsampling.layers" in name and name.endswith(".bias"):
            if len(data_torch.shape) == 1:
                data_torch = data_torch.reshape(1, -1, 1, 1)

        if "pointwise_conv" in name and name.endswith(".bias"):
            if len(data_torch.shape) == 1:
                data_torch = data_torch.reshape(1, -1, 1, 1)

        for mapped_name, tensor in super().modify_tensors(data_torch, name, bid):
            if name.startswith("sound_projection.") and mapped_name.startswith("mm.model.mlp."):
                mapped_name = mapped_name.replace("mm.model.mlp.", "mm.a.mlp.")
            yield mapped_name, tensor


@ModelBase.register("NemotronH_Omni_Reasoning_V3")
class NemotronOmniVLModel(NemotronNanoV2VLModel):
    _omni_tensor_prefixes = (
        "vision_model.embeddings.",
        "vision_model.encoder.layer.",
        "vision_projector.mlp1.",
    )

    def get_vision_config(self) -> dict[str, Any] | None:
        vision_config = dict(self.global_config["vision_config"])
        vision_config["intermediate_size"] = int(vision_config["hidden_size"] * vision_config["mlp_ratio"])
        vision_config["image_size"] = 512  # warmup only; the full position table is resized at runtime
        return vision_config

    def set_gguf_parameters(self):
        preprocessor_config = self.preprocessor_config

        preprocessor_config["image_mean"] = preprocessor_config["norm_mean"]
        preprocessor_config["image_std"] = preprocessor_config["norm_std"]

        MmprojModel.set_gguf_parameters(self)
        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_OMNI)
        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
        self.gguf_writer.add_vision_use_gelu(True)
        downsample_ratio = self.global_config.get("downsample_ratio", 0.5)
        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
        patch_area = self.hparams["patch_size"] ** 2
        self.gguf_writer.add_vision_min_pixels(preprocessor_config["min_num_patches"] * patch_area)
        self.gguf_writer.add_vision_max_pixels(preprocessor_config["max_num_patches"] * patch_area)

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        name, gen = item

        # Transformers can rename RADIO and projector tensors in Omni checkpoints.
        # Accept both checkpoint layouts.
        if name == "vision_model.summary_idxs":
            # This buffer is only used by RadioModel.summary. The multimodal
            # projector consumes RadioModel.features for both images and video.
            return None
        if name.startswith(cls._omni_tensor_prefixes):
            return name, gen

        if "radio_model.model.patch_generator.video_embedder" in name or name.startswith("vision_projector.vision_final_layernorm."):
            return MmprojModel.filter_tensors((name, gen))

        return super().filter_tensors(item)

    def tensor_force_quant(self, name, new_name, bid, n_dims):
        if new_name in ("v.position_embd.weight", "v.class_embd"):
            return gguf.GGMLQuantizationType.F32
        return super().tensor_force_quant(name, new_name, bid, n_dims)

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if name.startswith(self._omni_tensor_prefixes):
            if name == "vision_model.embeddings.video_patch_projection.weight":
                data_torch = data_torch.reshape(
                    self.hparams["hidden_size"],
                    6,
                    self.hparams["patch_size"],
                    self.hparams["patch_size"],
                )
                yield gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch
                return

            if name.startswith("vision_projector.mlp1."):
                projector_mapping = {
                    "vision_projector.mlp1.norm.weight": "mlp1.0.weight",
                    "vision_projector.mlp1.linear1.weight": "mlp1.1.weight",
                    "vision_projector.mlp1.linear2.weight": "mlp1.3.weight",
                }
                name = projector_mapping[name]
            elif name == "vision_model.embeddings.patch_projection.weight":
                data_torch = data_torch.reshape(
                    self.hparams["hidden_size"],
                    3,
                    self.hparams["patch_size"],
                    self.hparams["patch_size"],
                )
            elif name == "vision_model.embeddings.position_embedding":
                data_torch = data_torch.squeeze(0)
                name += ".weight"
            elif ".layer_scale" in name and name.endswith(".lambda1"):
                name = name.removesuffix("lambda1") + "weight"

            yield from MmprojModel.modify_tensors(self, data_torch, name, bid)
            return

        if "patch_generator.pos_embed" in name:
            # Keep the full table for runtime interpolation.
            yield from MmprojModel.modify_tensors(self, data_torch, name, bid)
            return

        if "patch_generator.video_embedder" in name:
            patch_size = self.hparams["patch_size"]
            n_embd = self.hparams["hidden_size"]
            temporal_patch_size = self.global_config["video_temporal_patch_size"]
            if temporal_patch_size != 2:
                raise ValueError("Current implementation only supports a temporal patch size of 2")
            data_torch = data_torch.reshape(n_embd, temporal_patch_size * 3, patch_size, patch_size)
            new_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1"
            yield new_name, data_torch
            return

        yield from super().modify_tensors(data_torch, name, bid)


@ModelBase.register("NemotronForCausalLM")
@ModelBase.example("nvidia/Minitron-4B-Base")
class NemotronModel(TextModel):
    model_arch = gguf.MODEL_ARCH.NEMOTRON

    def set_vocab(self):
        if (self.dir_model / "tokenizer.model").exists():
            self._set_vocab_sentencepiece()
            self.gguf_writer.add_pad_token_id(0)
            self.gguf_writer.add_unk_token_id(1)
        else:
            # SPM-style BPE (Metaspace + byte fallback): mark <0xXX> tokens as BYTE so they decode to raw bytes
            tokens, toktypes, tokpre = self.get_vocab_base()
            toktypes = [gguf.TokenType.BYTE if re.fullmatch(r"<0x[0-9A-F]{2}>", tok) else typ for tok, typ in zip(tokens, toktypes)]
            self.gguf_writer.add_tokenizer_model("gpt2")
            self.gguf_writer.add_tokenizer_pre(tokpre)
            self.gguf_writer.add_token_list(tokens)
            self.gguf_writer.add_token_types(toktypes)

            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
            special_vocab.add_to_gguf(self.gguf_writer)

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        hparams = self.hparams
        self.gguf_writer.add_vocab_size(hparams["vocab_size"])

        f_norm_eps = self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon", "norm_eps"])
        self.gguf_writer.add_layer_norm_eps(f_norm_eps)

        # * Partial RoPE
        rot_pct = self.rope_parameters["partial_rotary_factor"]
        n_embd = self.find_hparam(["hidden_size", "n_embd"])
        n_head = self.find_hparam(["num_attention_heads", "n_head"])
        # Use explicit head_dim when available; fallback keeps standard Nemotron working.
        # Without this, models where n_head * head_dim != n_embd (e.g. 48 heads, head_dim=128) produce the wrong rope dimension count.
        head_dim = self.hparams.get("head_dim") or n_embd // n_head
        self.gguf_writer.add_rope_dimension_count(int(rot_pct * head_dim))

        # * RopeScaling for Nemotron
        factor = self.hparams.get("factor") or self.rope_parameters.get("factor")
        if factor is None:
            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
        else:
            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
            self.gguf_writer.add_rope_scaling_factor(factor)

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        # * Adding +1 to LayerNorm's weights here to implement layernorm1p w/o changing anything on the GGML engine side
        #   model.layers.{l}.input_layernorm.weight
        #   model.layers.{l}.post_attention_layernorm.weight
        #   model.norm.weight
        if name.endswith("norm.weight"):
            data_torch = data_torch + 1

        yield from super().modify_tensors(data_torch, name, bid)


@ModelBase.register("NemotronHForCausalLM")
@ModelBase.example("nvidia/Nemotron-H-8B-Base-8K")
class NemotronHModel(GraniteHybridModel):
    """Hybrid mamba2/attention model from NVIDIA"""
    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
    is_moe: bool = False
    supports_mtp_export = True
    _experts: list[dict[str, Tensor]] | None = None

    _SSM_LAYER_TYPES = {"mamba", "linear_attention"}
    _ATTN_LAYER_TYPES = {"attention", "full_attention"}
    _MLP_LAYER_TYPES = {"moe"}

    def __init__(self, *args, **kwargs):
        # We have to determine the correct model architecture (MoE vs non-MoE) before
        # calling the parent __init__. This is because the parent constructor
        # uses self.model_arch to build the tensor name map, and all MoE-specific
        # mappings would be missed if it were called with the default non-MoE arch.
        hparams = kwargs.pop("hparams", None)
        if hparams is None:
            hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
        llm_config = {**hparams, **hparams.get("text_config", {})}

        has_moe_params = "num_experts_per_tok" in llm_config
        layers_block_type = llm_config.get("layers_block_type")

        if has_moe_params:
            self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
            self.is_moe = True
        if layers_block_type is not None:
            hparams["num_hidden_layers"] = len(layers_block_type)

        super().__init__(*args, hparams=hparams, **kwargs)

        # Save the top-level head_dim for later
        self.head_dim = self.hparams.get("head_dim", self.hparams.get("attention_head_dim"))
        assert self.head_dim is not None, "Could not find the attention head dim in config"

        # Don't use expand to calculate d_inner
        self.d_inner = self.find_hparam(["num_heads"]) * self.d_model

        # Update the ssm / attn / mlp layers
        # M: Mamba2, *: Attention, -: MLP
        # MoE:
        # M: Mamba2, *: Attention, E: Expert
        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
        if pattern is None:
            self._ssm_layers = []
            self._mlp_layers = []
        elif isinstance(pattern, str):
            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "M"]
            self._mlp_layers = [i for i, val in enumerate(pattern) if val == ("E" if self.is_moe else "-")]
        else:
            self._ssm_layers = [i for i, val in enumerate(pattern) if val in self._SSM_LAYER_TYPES]
            self._mlp_layers = [i for i, val in enumerate(pattern) if val in self._MLP_LAYER_TYPES]

        # `--no-mtp` drops it entirely; `--mtp` exports only the MTP head
        self._mtp_bid: int | None = None
        if self.is_moe and not self.no_mtp:
            n_nextn = self.hparams.get("num_nextn_predict_layers", 0) or 0
            if n_nextn > 0:
                assert n_nextn == 1, (
                    "NemotronH MTP conversion currently supports num_nextn_predict_layers == 1"
                )
                self._mtp_bid = self.block_count
                self.block_count += 1
                # The folded MTP block carries both an attention sub-layer and a
                # MoE sub-layer, so register it as both so the per-layer metadata arrays cover it
                self._attn_layers.append(self._mtp_bid)
                self._mlp_layers.append(self._mtp_bid)
                self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)

        if self.mtp_only and self._mtp_bid is None:
            raise ValueError("--mtp was requested, but this model does not contain a supported MTP head")

    def get_attn_layers(self):
        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
        if pattern is None:
            return []
        assert len(pattern) == self.block_count, f"Mismatch between pattern ({len(pattern)}) and block_count ({self.block_count})!"
        if isinstance(pattern, str):
            return [i for i, val in enumerate(pattern) if val == "*"]

        return [i for i, val in enumerate(pattern) if val in self._ATTN_LAYER_TYPES]

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        name, gen = item
        if name.startswith("language_model."):
            name = name[len("language_model."):]
        if name.startswith("mtp."):
            # --no-mtp: drop the MTP head entirely
            if cls.no_mtp:
                return None
        elif cls.mtp_only:
            # --mtp: export the MTP head plus the tensors it shares with the target model
            # Include lm_head scale sidecars so NVFP4 packing sees them.
            keep = name in (
                "backbone.embeddings.weight",
                "backbone.norm_f.weight",
                "lm_head.weight",
                "lm_head.weight_scale",
                "lm_head.weight_scale_2",
                "lm_head.weight_scale_inv",
                "lm_head.input_scale",
                "lm_head.input_global_scale",
                "lm_head.weight_global_scale",
                "lm_head.weight_packed",
            )
            if not keep:
                return None
        # PEFT names adapter tensors using model.layers.*, while Nemotron-H checkpoints
        # and the GGUF tensor map use backbone.layers.*
        if name.startswith("model.layers.") and ".mixer." in name:
            name = name.replace("model.layers.", "backbone.layers.", 1)
        return super().filter_tensors((name, gen))

    def prepare_metadata(self, vocab_only: bool):
        from_dir = self.fname_out.is_dir()
        super().prepare_metadata(vocab_only=vocab_only)

        if not self.mtp_only or not from_dir:
            return
        output_type: str = self.ftype.name.partition("_")[2]
        fname_default: str = gguf.naming_convention(
            self.metadata.name, self.metadata.basename, self.metadata.finetune,
            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"

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

        head_dim = self.head_dim
        if head_dim is None:
            raise ValueError("Could not find the attention head dim in config")
        self.gguf_writer.add_key_length(head_dim)
        self.gguf_writer.add_value_length(head_dim)

        # Set feed_forward_length
        # NOTE: This will trigger an override warning. This is preferable to
        #   duplicating all the parent logic
        if not self.is_moe:
            n_ff = self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"])
            self.gguf_writer.add_feed_forward_length([
                n_ff if i in self._mlp_layers else 0 for i in range(self.block_count)
            ])
        else:
            moe_intermediate_size = self.hparams["moe_intermediate_size"]
            self.gguf_writer.add_feed_forward_length([
                moe_intermediate_size if i in self._mlp_layers else 0 for i in range(self.block_count)
            ])
            self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
            self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
            self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
            self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
            self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
            self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
            self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
            self.gguf_writer.add_expert_group_count(self.hparams["n_group"])

            # number of experts used per token (top-k)
            if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
                self.gguf_writer.add_expert_used_count(n_experts_used)

            if (latent_size := self.hparams.get("moe_latent_size")) is not None:
                self.gguf_writer.add_moe_latent_size(latent_size)

        # MTP head: number of trailing NextN blocks
        if self._mtp_bid is not None:
            self.gguf_writer.add_nextn_predict_layers(self.hparams["num_nextn_predict_layers"])

    def set_vocab(self):
        # The NemotronH config uses pattern characters (e.g. '-') that may not
        # be supported by the installed transformers version. AutoTokenizer
        # internally calls AutoConfig which triggers this parsing failure.
        # Using trust_remote_code=True to load the model's own config class.
        tokens: list[str] = []
        toktypes: list[int] = []

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)

        # Pad vocab size (from Mamba2Model/GraniteHybridModel)
        self.hparams["pad_vocab_size_multiple"] = 8 # Setting this here since GraniteHybridModel.set_vocab() isn't being invoked now.
        # From Mamba2Model.set_vocab():
        vocab_size = self.hparams["vocab_size"]
        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
        # ref: https://stackoverflow.com/a/17511341/22827863
        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
        self.hparams["vocab_size"] = vocab_size

        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]

        tokpre = self.get_vocab_base_pre(tokenizer)

        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]

        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]

        for i in range(vocab_size):
            if i not in reverse_vocab:
                tokens.append(f"[PAD{i}]")
                toktypes.append(gguf.TokenType.UNUSED)
            else:
                token: str = reverse_vocab[i]
                if token in added_vocab:
                    if not added_tokens_decoder[i].normalized:
                        previous_token = token
                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
                        if previous_token != token:
                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
                        toktypes.append(gguf.TokenType.CONTROL)
                    else:
                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
                        toktypes.append(gguf.TokenType.USER_DEFINED)
                else:
                    toktypes.append(gguf.TokenType.NORMAL)
                tokens.append(token)

        # From TextModel.set_vocab_gpt2():
        self.gguf_writer.add_tokenizer_model("gpt2")
        self.gguf_writer.add_tokenizer_pre(tokpre)
        self.gguf_writer.add_token_list(tokens)
        self.gguf_writer.add_token_types(toktypes)

        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        special_vocab.add_to_gguf(self.gguf_writer)

    _MTP_SPECIAL_RENAMES = {
        "mtp.layers.0.enorm.weight":           "model.layers.{bid}.enorm.weight",
        "mtp.layers.0.hnorm.weight":           "model.layers.{bid}.hnorm.weight",
        "mtp.layers.0.eh_proj.weight":         "model.layers.{bid}.eh_proj.weight",
        "mtp.layers.1.norm.weight":            "model.layers.{bid}.post_attention_layernorm.weight",
        "mtp.layers.1.final_layernorm.weight": "model.layers.{bid}.shared_head.norm.weight",
    }

    def _generate_nvfp4_tensors(self):
        if self._mtp_bid is not None:
            for name in list(self.model_tensors):
                if name.startswith("mtp.layers.1.mixer"):
                    new_name = f"backbone.layers.{self._mtp_bid}.{name.split('.', 3)[3]}"
                    self.model_tensors[new_name] = self.model_tensors.pop(name)
        super()._generate_nvfp4_tensors()

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        #   mtp.layers.0: NextN input fusion + attention
        #   mtp.layers.1: MoE + final head norm
        if self._mtp_bid is not None and name.startswith(("mtp.layers.0.", "mtp.layers.1.")):
            suffix = name.split(".", 3)[3]
            bid = self._mtp_bid
            renamed = self._MTP_SPECIAL_RENAMES.get(name)
            name = renamed.format(bid=bid) if renamed else f"backbone.layers.{bid}.{suffix}"

        if self.is_moe and bid is not None:
            if name.endswith("mixer.gate.e_score_correction.bias"):
                yield from ModelBase.modify_tensors(self, data_torch, name, bid)
                return

            if name.endswith("mixer.dt_bias"):
                new_name = name.replace("dt_bias", "dt.bias")
                yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)
                return

            if name.endswith("mixer.conv1d.weight"):
                squeezed_data = data_torch.squeeze()
                yield from ModelBase.modify_tensors(self, squeezed_data, name, bid)
                return

            if name.endswith("mixer.A_log"):
                transformed_data = -torch.exp(data_torch)
                reshaped_data = transformed_data.squeeze().reshape(-1, 1)
                yield from ModelBase.modify_tensors(self, reshaped_data, name, bid)
                return

            if name.endswith("mixer.D"):
                reshaped_data = data_torch.squeeze().reshape(-1, 1)
                yield from ModelBase.modify_tensors(self, reshaped_data, name, bid)
                return

            if name.endswith("mixer.norm.weight"):
                reshaped_data = data_torch.reshape(self.n_group, -1)
                yield from ModelBase.modify_tensors(self, reshaped_data, name, bid)
                return

            if name.find("mixer.experts") != -1:
                n_experts = self.hparams["n_routed_experts"]
                assert bid is not None

                if self._experts is None:
                    self._experts = [{} for _ in range(self.block_count)]

                self._experts[bid][name] = data_torch

                if len(self._experts[bid]) >= n_experts * 2:
                    # merge the experts into a single tensor
                    for w_name in ["down_proj", "up_proj"]:
                        datas: list[Tensor] = []

                        for xid in range(n_experts):
                            ename = f"backbone.layers.{bid}.mixer.experts.{xid}.{w_name}.weight"
                            datas.append(self._experts[bid][ename])
                            del self._experts[bid][ename]

                        data_torch = torch.stack(datas, dim=0)
                        merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"

                        yield from ModelBase.modify_tensors(self, data_torch, merged_name, bid)
                    return
                else:
                    return

        yield from super().modify_tensors(data_torch, name, bid)

    def prepare_tensors(self):
        super().prepare_tensors()

        if self._experts is not None:
            # flatten `list[dict[str, Tensor]]` into `list[str]`
            experts = [k for d in self._experts for k in d.keys()]
            if len(experts) > 0:
                raise ValueError(f"Unprocessed experts: {experts}")


@ModelBase.register("NemotronHPuzzleForCausalLM")
@ModelBase.example("nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16")
class NemotronHPuzzleModel(NemotronHModel):
    """NVIDIA Puzzle: NemotronH with a per-block MoE config (block_configs).

    The checkpoint also ships an MTP draft head (mtp.safetensors). It is skipped
    here: there is no Puzzle MTP inference path in tree, and the head is laid out
    by mtp_block_configs rather than the mtp.layers.* form NemotronHModel maps."""

    model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
    is_moe: bool = True
    supports_mtp_export = False

    def __init__(self, dir_model: "Path", *args, **kwargs):
        hparams = dict(kwargs.pop("hparams", None) or ModelBase.load_hparams(dir_model, self.is_mistral_format))

        self.block_configs: list[dict] = hparams["block_configs"]
        self.n_layer_trunk = len(self.block_configs)

        # block_configs carries the per-block MoE shape, and is the authority on the
        # block pattern too: the layers_block_type the HF config wrapper computes is
        # not sized to it.
        hparams["num_hidden_layers"] = self.n_layer_trunk
        hparams["layers_block_type"] = [bc["block_type"] for bc in self.block_configs]

        self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE

        # Bypass NemotronHModel.__init__: it assumes a flat num_experts_per_tok /
        # moe_intermediate_size and a layers_block_type sized to block_count, neither
        # of which hold for Puzzle's per-block config.
        GraniteHybridModel.__init__(self, dir_model, *args, hparams=hparams, **kwargs)

        self.head_dim = self.find_hparam(["head_dim", "attention_head_dim"])
        self.d_inner = self.find_hparam(["num_heads"]) * self.d_model

        # NemotronHModel.__init__ folds an MTP block into block_count when the
        # config carries num_nextn_predict_layers; Puzzle's config does, but its
        # head has a different layout and no inference path, so stay opted out.
        self._mtp_bid = None

    def set_gguf_parameters(self):
        GraniteHybridModel.set_gguf_parameters(self)

        head_dim = self.head_dim
        if head_dim is None:
            raise ValueError("Could not find the attention head dim in config")
        self.gguf_writer.add_key_length(head_dim)
        self.gguf_writer.add_value_length(head_dim)

        ffn_lengths = [bc.get("moe_intermediate_size") or 0 for bc in self.block_configs]
        experts_used = [bc.get("num_experts_per_tok") or 0 for bc in self.block_configs]

        self.gguf_writer.add_feed_forward_length(ffn_lengths)
        self.gguf_writer.add_expert_feed_forward_length(ffn_lengths)
        self.gguf_writer.add_expert_used_count(experts_used)

        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
        self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
        self.gguf_writer.add_expert_group_count(self.hparams["n_group"])
        self.gguf_writer.add_moe_latent_size(self.hparams["moe_latent_size"])

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        # The official BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16)
        # names the trunk "model.*" (model.layers.*, model.embeddings, model.norm_f)
        # where the original release used the NemotronH-style "backbone.*", and spells
        # the router bias "e_score_correction_bias" instead of "e_score_correction.bias";
        # normalize so both convert identically.
        if name.startswith("model."):
            name = "backbone." + name[len("model."):]
        if name.endswith("mixer.gate.e_score_correction_bias"):
            name = name[: -len("e_score_correction_bias")] + "e_score_correction.bias"

        yield from super().modify_tensors(data_torch, name, bid)

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        # Drop the MTP head unconditionally; see the class docstring.
        if item[0].startswith("mtp."):
            return None
        return super().filter_tensors(item)