Buckets:

gijl
/

lv-storage

Files

xet

gijl/lv-storage / llama.cpp /conversion /qwen4exp.py

gijl

15 days ago

download

raw

12.4 kB

from __future__ import annotations

from typing import Iterable, cast

import torch
from torch import Tensor

import gguf
import numpy as np

from .base import ModelBase
from .qwen import _LinearAttentionVReorderBase, _Qwen35MRopeMixin
from .qwen3vl import Qwen3VLVisionModel


@ModelBase.register("Qwen4ExpForConditionalGeneration", "Qwen4ExpForCausalLM")
@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
    """Qwen3.8-Flash-Next.
    Shares the Qwen3.5 gated delta net and interleaved mrope, and adds three things:
    hyper-connections in place of every layer norm, QSA sparse attention on the full
    attention layers, and PLE n-gram hash embeddings on a single layer.
    """

    model_arch = gguf.MODEL_ARCH.QWEN4EXP

    # the MTP head: one full-attention QSA block after the trunk, fed by the trunk's hc-wide residual
    supports_mtp_export = True

    # MTP tensors the shared Qwen remapper does not know
    _MTP_EXTRA = {
        "fc_embedding":           "nextn_fc_embedding",
        "fc_hidden":              "nextn_fc_hidden",
        "hyper_connection_mixer": "nextn_hc_head",
    }

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        # only the shard names, so the table itself is never held
        self._ple_shards: dict[int, str] = {}
        self._ple_row_dim: int | None = None
        self._mtp_fc: dict[str, Tensor] = {}
        # halves of the merged weights, keyed by (bid, merged tensor), see _merge_shared_input
        self._merge_parts: dict[tuple[int, gguf.MODEL_TENSOR], dict[int, Tensor]] = {}

    @classmethod
    def filter_tensors(cls, item):
        name, gen = item
        part = name.split(".")[1] if name.startswith("mtp.") else None
        if part in cls._MTP_EXTRA:
            if cls.no_mtp:
                return None
            assert cls._original_block_count is not None
            rest = name.split(".", 2)[2]
            return f"model.layers.{cls._original_block_count}.{cls._MTP_EXTRA[part]}.{rest}", gen
        return super().filter_tensors(item)

    def _read_hash_constants(self, suffix: str) -> list[int]:
        """Read an int64 PLE constant straight from the checkpoint.
        prepare_tensors() casts every non-float dtype to float32 before
        modify_tensors() sees it (base.py), which would silently round these
        45-bit multipliers. Reading the lazy tensor here bypasses that.
        """
        for name, gen in self.model_tensors.items():
            if name.endswith(suffix):
                t = gen()
                if t.dtype != torch.int64:
                    t = t.to(torch.int64)
                return [int(x) for x in t.tolist()]
        raise ValueError(f"PLE constant {suffix!r} missing from the checkpoint")

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        hp = self.hparams

        self.gguf_writer.add_hyper_connection_count(hp["hc_count"])
        self.gguf_writer.add_hyper_connection_low_rank(hp["hc_lowrank"])

        n_layer = hp["num_hidden_layers"]
        self.gguf_writer.add_indexer_head_count(hp["indexer_n_heads"])
        self.gguf_writer.add_indexer_key_length(hp["indexer_head_dim"])
        self.gguf_writer.add_indexer_top_k(hp["indexer_budget"])
        ratio = hp["indexer_compress_ratio"]
        layer_types = hp["layer_types"]
        # the MTP block is a full-attention QSA layer too
        self.gguf_writer.add_attention_compress_ratios(
            [ratio if layer_types[i] == "full_attention" else 0 for i in range(n_layer)]
            + [ratio] * (self.block_count - n_layer)
        )

        # ple_layer_ids is 1-based in the HF config; empty means no n-gram table,
        # so emit no PLE keys rather than optional ones
        # the MTP head never reads PLE, so an MTP-only file carries none of it
        ple_layers = [i - 1 for i in hp["ple_layer_ids"]]
        if not ple_layers or self.mtp_only:
            return
        self.gguf_writer.add_ple_layers(ple_layers)
        self.gguf_writer.add_ple_ngram_size(hp["ngram_size"])
        self.gguf_writer.add_ple_heads_per_ngram(hp["heads_per_ngram"])
        self.gguf_writer.add_ple_conv_kernel(hp["ple_conv_kernel_size"])
        self.gguf_writer.add_ple_eos_token_id(self._eos_token_id())
        # an image is decoded as an embeddings-only batch, so the graph has no placeholder
        # ids to hash; carry the id and let it stand in for those positions
        _img = self._image_token_id()
        if _img is not None:
            self.gguf_writer.add_ple_image_token_id(int(_img))
        if self._ple_row_dim is not None:
            self.gguf_writer.add_embedding_length_per_layer_input(self._ple_row_dim)

        self.gguf_writer.add_ple_layer_multipliers(
            self._read_hash_constants("ple_embedding.layer_multipliers"))
        self.gguf_writer.add_ple_head_offsets(
            self._read_hash_constants("ple_embedding.ngram_heads_offsets"))
        self.gguf_writer.add_ple_head_vocab_sizes(
            self._read_hash_constants("ple_embedding.ngram_heads_vocab_sizes"))

    def _image_token_id(self) -> int | None:
        img = self.hparams.get("image_token_id")
        return None if img is None else int(img)

    def _eos_token_id(self) -> int:
        eos = self.hparams.get("eos_token_id")
        if isinstance(eos, list):
            # the PLE hash resets n-grams on the primary EOS
            return int(eos[-1])
        if eos is None:
            raise ValueError("eos_token_id is required: the PLE hash resets its n-grams on it")
        return int(eos)

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        # int64 hash constants must stay exact; 1-D tensors force F32, so use KV
        if name.endswith("ple_embedding.layer_multipliers"):
            self._ple_multipliers = [int(x) for x in data_torch.tolist()]
            return
        if name.endswith("ple_embedding.ngram_heads_offsets"):
            self._ple_head_offsets = [int(x) for x in data_torch.tolist()]
            return
        if name.endswith("ple_embedding.ngram_heads_vocab_sizes"):
            self._ple_head_vocab_sizes = [int(x) for x in data_torch.tolist()]
            return

        if ".ngram_embedding.shard_" in name:
            yield from self._place_ple_shard(data_torch, name)
            return

        # eh_proj([e ; h_s]) = fc_embedding(e) + fc_hidden(h_s) for every hc stream s
        if name.endswith((".nextn_fc_embedding.weight", ".nextn_fc_hidden.weight")):
            self._mtp_fc[name.rsplit(".", 2)[1]] = data_torch
            if len(self._mtp_fc) < 2:
                return
            eh = torch.cat([self._mtp_fc.pop("nextn_fc_embedding"), self._mtp_fc.pop("nextn_fc_hidden")], dim=1)
            yield (self.format_tensor_name(gguf.MODEL_TENSOR.NEXTN_EH_PROJ, bid, ".weight"), eh)
            return

        # one projection feeds indexer q and k; split it, as minimax-m3 does
        if ".indexer.index_qk_proj.weight" in name:
            n_q = self.hparams["indexer_n_heads"] * self.hparams["indexer_head_dim"]
            yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_PROJ, bid, ".weight"), data_torch[:n_q])
            yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_K_PROJ, bid, ".weight"), data_torch[n_q:])
            return

        # Gemma zero-centred gammas the inherited norm.weight rule misses
        if name.endswith((".ple.norm_key.weight", ".ple.norm_query.weight", ".ple.norm_conv.weight",
                          ".indexer.q_layernorm.weight", ".indexer.k_layernorm.weight")):
            yield (self.map_tensor_name(name), data_torch + 1)
            return

        if name.endswith(".ple.conv1d.weight"):
            yield (self.map_tensor_name(name), data_torch.squeeze())
            return

        for new_name, data in super().modify_tensors(data_torch, name, bid):
            yield from self._merge_shared_input(new_name, data, bid)

    # matmuls that read the same input, written as one weight [a | b] so the runtime does one matmul:
    # (merged tensor, (a, b), linear-attention layers only)
    _SHARED_INPUT_MERGES = (
        (gguf.MODEL_TENSOR.HC_ATTN_DOWN_INJECT, (gguf.MODEL_TENSOR.HC_ATTN_DOWN, gguf.MODEL_TENSOR.HC_ATTN_INJECT), False),
        (gguf.MODEL_TENSOR.HC_FFN_DOWN_INJECT,  (gguf.MODEL_TENSOR.HC_FFN_DOWN,  gguf.MODEL_TENSOR.HC_FFN_INJECT),  False),
        (gguf.MODEL_TENSOR.ATTN_QKV_GATE,       (gguf.MODEL_TENSOR.ATTN_QKV,     gguf.MODEL_TENSOR.ATTN_GATE),      True),
        (gguf.MODEL_TENSOR.SSM_BETA_ALPHA_CAT,  (gguf.MODEL_TENSOR.SSM_BETA,     gguf.MODEL_TENSOR.SSM_ALPHA),      True),
    )

    # caches one half of a merged weight until the other half arrives, then yields [a | b]
    def _merge_shared_input(self, new_name: str, data: Tensor, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        # a pre-quantized checkpoint carries per-tensor scales, which one merged weight cannot hold
        if bid is None or self.hparams.get("quantization_config"):
            yield (new_name, data)
            return

        layer_types = self.hparams.get("layer_types") or []
        is_linear = bid < len(layer_types) and layer_types[bid] == "linear_attention"

        match = next(((merged, parts.index(part)) for merged, parts, linear_only in self._SHARED_INPUT_MERGES
                      for part in parts
                      if (is_linear or not linear_only) and self.match_model_tensor_name(new_name, part, bid)), None)
        if match is None:
            yield (new_name, data)
            return

        merged, idx = match
        halves = self._merge_parts.setdefault((bid, merged), {})
        halves[idx] = data
        if len(halves) == 2:
            del self._merge_parts[(bid, merged)]
            yield (self.format_tensor_name(merged, bid), torch.cat([halves[0], halves[1]], dim=0))

    # the shards concatenate into a tensor of well over 100 GB
    # use LazyChunkedTensor here, a single shard resident at a time
    def _place_ple_shard(self, data_torch: Tensor, name: str) -> Iterable[tuple[str, Tensor]]:

        idx = int(name.rpartition(".shard_")[2].partition(".")[0])
        n_parts = self.hparams["split_ngram_parts"]

        self._ple_shards[idx] = name
        self._ple_row_dim = int(data_torch.shape[-1])

        if len(self._ple_shards) < n_parts:
            return []

        # the checkpoint may yield the shards in any order, the row order is by index
        shards = [self._ple_shards[i] for i in sorted(self._ple_shards)]
        rows = 0
        for shard in shards:
            shape = self.model_tensors[shard]().shape
            if int(shape[-1]) != self._ple_row_dim:
                raise ValueError(
                    f"PLE shard {shard} has row dim {int(shape[-1])}, expected {self._ple_row_dim}")
            rows += int(shape[0])

        table = gguf.LazyChunkedTensor(
            [self._load_ple_shard(shard) for shard in shards],
            shape=(rows, self._ple_row_dim),
            dtype=np.float32,
        )
        gguf_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD]
        return [(gguf_name + ".weight", cast(Tensor, table))]

    def _load_ple_shard(self, name: str):
        def load() -> np.ndarray:
            from .base import LazyTorchTensor

            # a fresh lazy tensor every call, or to_eager() memoizes every shard
            eager = LazyTorchTensor.to_eager(self.model_tensors[name]())
            return eager.to(torch.float32).contiguous().numpy()
        return load

    def prepare_tensors(self):
        super().prepare_tensors()
        if self._mtp_fc:
            raise ValueError(f"MTP projection missing its other half: {sorted(self._mtp_fc)}")
        if self._merge_parts:
            raise ValueError(f"merged weights missing a half: {sorted((bid, t.name) for bid, t in self._merge_parts)}")
        n_parts = self.hparams.get("split_ngram_parts", 0)
        if self._ple_shards and len(self._ple_shards) != n_parts:
            raise ValueError(
                f"got {len(self._ple_shards)} PLE embedding shards, expected {n_parts}"
            )


@ModelBase.register("Qwen4ExpForConditionalGeneration")
@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
class Qwen4ExpVisionModel(Qwen3VLVisionModel):
    """The vision tower is an unmodified Qwen3-VL ViT."""

Xet Storage Details

Size:: 12.4 kB
Xet hash:: a366b8ca7f72b7f16372344c226b961f7e9e219899e981b828139b9f0a9c6811

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.

	from __future__ import annotations

	from typing import Iterable, cast

	import torch
	from torch import Tensor

	import gguf
	import numpy as np

	from .base import ModelBase
	from .qwen import _LinearAttentionVReorderBase, _Qwen35MRopeMixin
	from .qwen3vl import Qwen3VLVisionModel


	@ModelBase.register("Qwen4ExpForConditionalGeneration", "Qwen4ExpForCausalLM")
	@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
	class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
	"""Qwen3.8-Flash-Next.

	Shares the Qwen3.5 gated delta net and interleaved mrope, and adds three things:
	hyper-connections in place of every layer norm, QSA sparse attention on the full
	attention layers, and PLE n-gram hash embeddings on a single layer.
	"""

	model_arch = gguf.MODEL_ARCH.QWEN4EXP

	# the MTP head: one full-attention QSA block after the trunk, fed by the trunk's hc-wide residual
	supports_mtp_export = True

	# MTP tensors the shared Qwen remapper does not know
	_MTP_EXTRA = {
	"fc_embedding": "nextn_fc_embedding",
	"fc_hidden": "nextn_fc_hidden",
	"hyper_connection_mixer": "nextn_hc_head",
	}

	def __init__(self, args, *kwargs):
	super().__init__(args, *kwargs)
	# only the shard names, so the table itself is never held
	self._ple_shards: dict[int, str] = {}
	self._ple_row_dim: int \| None = None
	self._mtp_fc: dict[str, Tensor] = {}
	# halves of the merged weights, keyed by (bid, merged tensor), see _merge_shared_input
	self._merge_parts: dict[tuple[int, gguf.MODEL_TENSOR], dict[int, Tensor]] = {}

	@classmethod
	def filter_tensors(cls, item):
	name, gen = item
	part = name.split(".")[1] if name.startswith("mtp.") else None
	if part in cls._MTP_EXTRA:
	if cls.no_mtp:
	return None
	assert cls._original_block_count is not None
	rest = name.split(".", 2)[2]
	return f"model.layers.{cls._original_block_count}.{cls._MTP_EXTRA[part]}.{rest}", gen
	return super().filter_tensors(item)

	def _read_hash_constants(self, suffix: str) -> list[int]:
	"""Read an int64 PLE constant straight from the checkpoint.

	prepare_tensors() casts every non-float dtype to float32 before
	modify_tensors() sees it (base.py), which would silently round these
	45-bit multipliers. Reading the lazy tensor here bypasses that.
	"""
	for name, gen in self.model_tensors.items():
	if name.endswith(suffix):
	t = gen()
	if t.dtype != torch.int64:
	t = t.to(torch.int64)
	return [int(x) for x in t.tolist()]
	raise ValueError(f"PLE constant {suffix!r} missing from the checkpoint")

	def set_gguf_parameters(self):
	super().set_gguf_parameters()
	hp = self.hparams

	self.gguf_writer.add_hyper_connection_count(hp["hc_count"])
	self.gguf_writer.add_hyper_connection_low_rank(hp["hc_lowrank"])

	n_layer = hp["num_hidden_layers"]
	self.gguf_writer.add_indexer_head_count(hp["indexer_n_heads"])
	self.gguf_writer.add_indexer_key_length(hp["indexer_head_dim"])
	self.gguf_writer.add_indexer_top_k(hp["indexer_budget"])
	ratio = hp["indexer_compress_ratio"]
	layer_types = hp["layer_types"]
	# the MTP block is a full-attention QSA layer too
	self.gguf_writer.add_attention_compress_ratios(
	[ratio if layer_types[i] == "full_attention" else 0 for i in range(n_layer)]
	+ [ratio] * (self.block_count - n_layer)
	)

	# ple_layer_ids is 1-based in the HF config; empty means no n-gram table,
	# so emit no PLE keys rather than optional ones
	# the MTP head never reads PLE, so an MTP-only file carries none of it
	ple_layers = [i - 1 for i in hp["ple_layer_ids"]]
	if not ple_layers or self.mtp_only:
	return
	self.gguf_writer.add_ple_layers(ple_layers)
	self.gguf_writer.add_ple_ngram_size(hp["ngram_size"])
	self.gguf_writer.add_ple_heads_per_ngram(hp["heads_per_ngram"])
	self.gguf_writer.add_ple_conv_kernel(hp["ple_conv_kernel_size"])
	self.gguf_writer.add_ple_eos_token_id(self._eos_token_id())
	# an image is decoded as an embeddings-only batch, so the graph has no placeholder
	# ids to hash; carry the id and let it stand in for those positions
	_img = self._image_token_id()
	if _img is not None:
	self.gguf_writer.add_ple_image_token_id(int(_img))
	if self._ple_row_dim is not None:
	self.gguf_writer.add_embedding_length_per_layer_input(self._ple_row_dim)

	self.gguf_writer.add_ple_layer_multipliers(
	self._read_hash_constants("ple_embedding.layer_multipliers"))
	self.gguf_writer.add_ple_head_offsets(
	self._read_hash_constants("ple_embedding.ngram_heads_offsets"))
	self.gguf_writer.add_ple_head_vocab_sizes(
	self._read_hash_constants("ple_embedding.ngram_heads_vocab_sizes"))

	def _image_token_id(self) -> int \| None:
	img = self.hparams.get("image_token_id")
	return None if img is None else int(img)

	def _eos_token_id(self) -> int:
	eos = self.hparams.get("eos_token_id")
	if isinstance(eos, list):
	# the PLE hash resets n-grams on the primary EOS
	return int(eos[-1])
	if eos is None:
	raise ValueError("eos_token_id is required: the PLE hash resets its n-grams on it")
	return int(eos)

	def modify_tensors(self, data_torch: Tensor, name: str, bid: int \| None) -> Iterable[tuple[str, Tensor]]:
	# int64 hash constants must stay exact; 1-D tensors force F32, so use KV
	if name.endswith("ple_embedding.layer_multipliers"):
	self._ple_multipliers = [int(x) for x in data_torch.tolist()]
	return
	if name.endswith("ple_embedding.ngram_heads_offsets"):
	self._ple_head_offsets = [int(x) for x in data_torch.tolist()]
	return
	if name.endswith("ple_embedding.ngram_heads_vocab_sizes"):
	self._ple_head_vocab_sizes = [int(x) for x in data_torch.tolist()]
	return

	if ".ngram_embedding.shard_" in name:
	yield from self._place_ple_shard(data_torch, name)
	return

	# eh_proj([e ; h_s]) = fc_embedding(e) + fc_hidden(h_s) for every hc stream s
	if name.endswith((".nextn_fc_embedding.weight", ".nextn_fc_hidden.weight")):
	self._mtp_fc[name.rsplit(".", 2)[1]] = data_torch
	if len(self._mtp_fc) < 2:
	return
	eh = torch.cat([self._mtp_fc.pop("nextn_fc_embedding"), self._mtp_fc.pop("nextn_fc_hidden")], dim=1)
	yield (self.format_tensor_name(gguf.MODEL_TENSOR.NEXTN_EH_PROJ, bid, ".weight"), eh)
	return

	# one projection feeds indexer q and k; split it, as minimax-m3 does
	if ".indexer.index_qk_proj.weight" in name:
	n_q = self.hparams["indexer_n_heads"] * self.hparams["indexer_head_dim"]
	yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_PROJ, bid, ".weight"), data_torch[:n_q])
	yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_K_PROJ, bid, ".weight"), data_torch[n_q:])
	return

	# Gemma zero-centred gammas the inherited norm.weight rule misses
	if name.endswith((".ple.norm_key.weight", ".ple.norm_query.weight", ".ple.norm_conv.weight",
	".indexer.q_layernorm.weight", ".indexer.k_layernorm.weight")):
	yield (self.map_tensor_name(name), data_torch + 1)
	return

	if name.endswith(".ple.conv1d.weight"):
	yield (self.map_tensor_name(name), data_torch.squeeze())
	return

	for new_name, data in super().modify_tensors(data_torch, name, bid):
	yield from self._merge_shared_input(new_name, data, bid)

	# matmuls that read the same input, written as one weight [a \| b] so the runtime does one matmul:
	# (merged tensor, (a, b), linear-attention layers only)
	_SHARED_INPUT_MERGES = (
	(gguf.MODEL_TENSOR.HC_ATTN_DOWN_INJECT, (gguf.MODEL_TENSOR.HC_ATTN_DOWN, gguf.MODEL_TENSOR.HC_ATTN_INJECT), False),
	(gguf.MODEL_TENSOR.HC_FFN_DOWN_INJECT, (gguf.MODEL_TENSOR.HC_FFN_DOWN, gguf.MODEL_TENSOR.HC_FFN_INJECT), False),
	(gguf.MODEL_TENSOR.ATTN_QKV_GATE, (gguf.MODEL_TENSOR.ATTN_QKV, gguf.MODEL_TENSOR.ATTN_GATE), True),
	(gguf.MODEL_TENSOR.SSM_BETA_ALPHA_CAT, (gguf.MODEL_TENSOR.SSM_BETA, gguf.MODEL_TENSOR.SSM_ALPHA), True),
	)

	# caches one half of a merged weight until the other half arrives, then yields [a \| b]
	def _merge_shared_input(self, new_name: str, data: Tensor, bid: int \| None) -> Iterable[tuple[str, Tensor]]:
	# a pre-quantized checkpoint carries per-tensor scales, which one merged weight cannot hold
	if bid is None or self.hparams.get("quantization_config"):
	yield (new_name, data)
	return

	layer_types = self.hparams.get("layer_types") or []
	is_linear = bid < len(layer_types) and layer_types[bid] == "linear_attention"

	match = next(((merged, parts.index(part)) for merged, parts, linear_only in self._SHARED_INPUT_MERGES
	for part in parts
	if (is_linear or not linear_only) and self.match_model_tensor_name(new_name, part, bid)), None)
	if match is None:
	yield (new_name, data)
	return

	merged, idx = match
	halves = self._merge_parts.setdefault((bid, merged), {})
	halves[idx] = data
	if len(halves) == 2:
	del self._merge_parts[(bid, merged)]
	yield (self.format_tensor_name(merged, bid), torch.cat([halves[0], halves[1]], dim=0))

	# the shards concatenate into a tensor of well over 100 GB
	# use LazyChunkedTensor here, a single shard resident at a time
	def _place_ple_shard(self, data_torch: Tensor, name: str) -> Iterable[tuple[str, Tensor]]:

	idx = int(name.rpartition(".shard_")[2].partition(".")[0])
	n_parts = self.hparams["split_ngram_parts"]

	self._ple_shards[idx] = name
	self._ple_row_dim = int(data_torch.shape[-1])

	if len(self._ple_shards) < n_parts:
	return []

	# the checkpoint may yield the shards in any order, the row order is by index
	shards = [self._ple_shards[i] for i in sorted(self._ple_shards)]
	rows = 0
	for shard in shards:
	shape = self.model_tensors[shard]().shape
	if int(shape[-1]) != self._ple_row_dim:
	raise ValueError(
	f"PLE shard {shard} has row dim {int(shape[-1])}, expected {self._ple_row_dim}")
	rows += int(shape[0])

	table = gguf.LazyChunkedTensor(
	[self._load_ple_shard(shard) for shard in shards],
	shape=(rows, self._ple_row_dim),
	dtype=np.float32,
	)
	gguf_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD]
	return [(gguf_name + ".weight", cast(Tensor, table))]

	def _load_ple_shard(self, name: str):
	def load() -> np.ndarray:
	from .base import LazyTorchTensor

	# a fresh lazy tensor every call, or to_eager() memoizes every shard
	eager = LazyTorchTensor.to_eager(self.model_tensors[name]())
	return eager.to(torch.float32).contiguous().numpy()
	return load

	def prepare_tensors(self):
	super().prepare_tensors()
	if self._mtp_fc:
	raise ValueError(f"MTP projection missing its other half: {sorted(self._mtp_fc)}")
	if self._merge_parts:
	raise ValueError(f"merged weights missing a half: {sorted((bid, t.name) for bid, t in self._merge_parts)}")
	n_parts = self.hparams.get("split_ngram_parts", 0)
	if self._ple_shards and len(self._ple_shards) != n_parts:
	raise ValueError(
	f"got {len(self._ple_shards)} PLE embedding shards, expected {n_parts}"
	)


	@ModelBase.register("Qwen4ExpForConditionalGeneration")
	@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
	class Qwen4ExpVisionModel(Qwen3VLVisionModel):
	"""The vision tower is an unmodified Qwen3-VL ViT."""