From 141dd55e53079a15851e1c8ec87e360d1d07876c Mon Sep 17 00:00:00 2001 From: Francis Couture-Harpin Date: Sun, 8 Sep 2024 20:01:13 -0400 Subject: [PATCH 1/2] convert : refactor rope_freqs generation This should also fix vocab-only conversion for Phi-3. --- convert_hf_to_gguf.py | 36 +++++++++++++++++++++--------------- convert_lora_to_gguf.py | 5 ++++- gguf-py/gguf/constants.py | 2 ++ 3 files changed, 27 insertions(+), 16 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 0a9bbc829..2dfc91446 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -15,6 +15,7 @@ from enum import IntEnum from pathlib import Path from hashlib import sha256 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast +from itertools import chain import math import numpy as np @@ -64,7 +65,6 @@ class Model: model_name: str | None metadata_override: Path | None dir_model_card: Path - is_lora: bool # subclasses should define this! model_arch: gguf.MODEL_ARCH @@ -72,7 +72,7 @@ class Model: def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, use_temp_file: bool = False, eager: bool = False, metadata_override: Path | None = None, model_name: str | None = None, - split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False): + split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): if type(self) is Model: raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") @@ -94,7 +94,6 @@ class Model: self.metadata_override = metadata_override self.model_name = model_name self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py - self.is_lora = is_lora # true if model is used inside convert_lora_to_gguf.py # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type if self.ftype == gguf.LlamaFileType.GUESSED: @@ -259,10 +258,14 @@ class Model: return False + # some models need extra generated tensors (like rope_freqs) + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + return () + def prepare_tensors(self): max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,") - for name, data_torch in self.get_tensors(): + for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()): # we don't need these if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): continue @@ -1590,7 +1593,7 @@ class LlamaModel(Model): return [(self.map_tensor_name(name), data_torch)] - def prepare_tensors(self): + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): if rope_scaling.get("rope_type", '').lower() == "llama3": base = self.hparams.get("rope_theta", 10000.0) @@ -1617,9 +1620,9 @@ class LlamaModel(Model): smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) rope_factors.append(1 / ((1 - smooth) / factor + smooth)) - if not self.is_lora: - self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) + def prepare_tensors(self): super().prepare_tensors() if self._experts is not None: @@ -2135,6 +2138,13 @@ class Phi3MiniModel(Model): self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"])) + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + n_embd = self.find_hparam(["hidden_size", "n_embd"]) + n_head = self.find_hparam(["num_attention_heads", "n_head"]) + max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"]) + orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"]) + rope_dims = n_embd // n_head + # write rope scaling for long context (128k) model rope_scaling = self.find_hparam(['rope_scaling'], True) if rope_scaling is None: @@ -2164,9 +2174,8 @@ class Phi3MiniModel(Model): if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') - if not self.is_lora: - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32)) @Model.register("PlamoForCausalLM") @@ -3915,7 +3924,7 @@ class ExaoneModel(Model): self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"]) - def prepare_tensors(self): + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): if rope_scaling.get("rope_type", '').lower() == "llama3": base = self.hparams.get("rope_theta", 10000.0) @@ -3942,10 +3951,7 @@ class ExaoneModel(Model): smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) rope_factors.append(1 / ((1 - smooth) / factor + smooth)) - if not self.is_lora: - self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) - - super().prepare_tensors() + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) ###### CONVERSION LOGIC ###### diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py index ddd347a2a..ef088034c 100755 --- a/convert_lora_to_gguf.py +++ b/convert_lora_to_gguf.py @@ -331,6 +331,10 @@ if __name__ == '__main__': self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha) super().set_gguf_parameters() + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + # Never add extra tensors (e.g. rope_freqs) for LoRA adapters + return () + def get_tensors(self) -> Iterator[tuple[str, Tensor]]: tensor_map: dict[str, PartialLoraTensor] = {} @@ -386,7 +390,6 @@ if __name__ == '__main__': dry_run=args.dry_run, dir_lora_model=dir_lora, lora_alpha=alpha, - is_lora=True, ) logger.info("Exporting model...") diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index c87d08782..988f207db 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -793,6 +793,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_QKV, MODEL_TENSOR.ATTN_Q, From e83d2707d3edae76e2d00408b0e0cfa39468d509 Mon Sep 17 00:00:00 2001 From: Francis Couture-Harpin Date: Mon, 16 Sep 2024 12:05:29 -0400 Subject: [PATCH 2/2] convert : adapt MiniCPM3 to separate rope_freqs insertion MiniCPM3's tokenizer is treated as a SentencePiece tokenizer to avoid having to run its custom Python code which mixes tokenization in the same file as tool calls. gguf-py : add long and short RoPE factors to tensor mappings Empty, but the key names are used to populate the mappings. --- convert_hf_to_gguf.py | 25 ++++++++++++------------- gguf-py/gguf/constants.py | 2 ++ gguf-py/gguf/tensor_mapping.py | 3 +++ 3 files changed, 17 insertions(+), 13 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index f333a567f..7c2c87e0b 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -1862,8 +1862,6 @@ class MiniCPM3Model(Model): def set_gguf_parameters(self): hparams = self.hparams - rope_dims = hparams["qk_rope_head_dim"] - self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) @@ -1879,24 +1877,25 @@ class MiniCPM3Model(Model): self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: rope_scaling = self.find_hparam(['rope_scaling'], True) - if rope_scaling is None: - return + if rope_scaling is not None: + rope_dims = self.hparams["qk_rope_head_dim"] - long_factors = rope_scaling.get('long_factor', None) - short_factors = rope_scaling.get('short_factor', None) + long_factors = rope_scaling.get('long_factor', None) + short_factors = rope_scaling.get('short_factor', None) - if long_factors is None or short_factors is None: - raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') + if long_factors is None or short_factors is None: + raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') - if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: - raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') + if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: + raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32)) def set_vocab(self): - self._set_vocab_llama_hf() + self._set_vocab_sentencepiece() def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: if n_kv_head is not None and n_head != n_kv_head: diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 0e7f86f0a..7aaf6745a 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -877,6 +877,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_Q_A, MODEL_TENSOR.ATTN_Q_B, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 2ebfa2b43..4639f2f9b 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -87,6 +87,9 @@ class TensorNameMap: "rope.freqs", # llama-pth "rotary_pos_emb.inv_freq", # chatglm ), + + MODEL_TENSOR.ROPE_FACTORS_LONG: (), + MODEL_TENSOR.ROPE_FACTORS_SHORT: (), } block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {