|
--- |
|
license: apache-2.0 |
|
tags: |
|
- mixtral |
|
- dense |
|
- mistral |
|
- expert |
|
--- |
|
|
|
# Unmixtraled 22B expert 3 |
|
|
|
> [!WARNING] |
|
> This model outputs gibberish as it was not trained under the dense configuration. Finetuning or merging is needed to make this model useful. |
|
|
|
This is a 22B Mistral model recycling weights from [mistral-community/Mixtral-8x22B-v0.1](https://huggingface.co/mistral-community/Mixtral-8x22B-v0.1). |
|
The model was adapted from a Mixtral architecture to a dense Mistral architecture with the same number of layers, attention heads and hidden dimensions. |
|
Embeddings, attention, layer norms and LM head weights were taken directly from the 8x22B model, all MLP weights were taken from expert 3. |
|
|
|
The following named weight correspondance was used: |
|
|
|
| Mistral weight | Mixtral weight | |
|
|----------------|----------------------------------| |
|
| `gate_proj` | `experts.3.w1` | |
|
| `down_proj` | `experts.3.w2` | |
|
| `up_proj` | `experts.3.w3` | |
|
|
|
## Unmixtraled models |
|
| Expert | Source | Wikitext perplexity | |
|
|--------|-----------------|---------------------| |
|
| [Unmixtraled-22B-v0.1-expert-0](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-0) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 0 MLPs | 696.6932983398438 | |
|
| [Unmixtraled-22B-v0.1-expert-1](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-1) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 1 MLPs | 6853.04248046875 | |
|
| [Unmixtraled-22B-v0.1-expert-2](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-2) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 2 MLPs | 4689.181640625 | |
|
| [**Unmixtraled-22B-v0.1-expert-3**](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-3) | **Mixtral 8x22B embed, attn, layernorm, lm_head + expert 3 MLPs** | **782.3755493164062** | |
|
| [Unmixtraled-22B-v0.1-expert-4](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-4) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 4 MLPs | 2844.943603515625 | |
|
| [Unmixtraled-22B-v0.1-expert-5](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-5) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 5 MLPs | 1099.32373046875 | |
|
| [Unmixtraled-22B-v0.1-expert-6](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-6) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 6 MLPs | 341.5309753417969 | |
|
| [Unmixtraled-22B-v0.1-expert-7](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-expert-7) | Mixtral 8x22B embed, attn, layernorm, lm_head + expert 7 MLPs | 2099.63818359375 | |
|
| [Unmixtraled-22B-v0.1-lerp](https://huggingface.co/thomasgauthier/Unmixtraled-22B-v0.1-lerp) | Mixtral 8x22B embed, attn, layernorm, lm_head + linear merge of expert 0-7 MLPs | 1873.9874267578125 | |
|
|
|
# Code |
|
|
|
The following code was used to extract the experts and construct the dense models: |
|
|
|
```python |
|
# pip install -U transformers huggingface_hub "git+https://github.com/arcee-ai/mergekit@7467108c05d56ef2bb4b8f33936d437dc448f7dd" |
|
|
|
import fnmatch |
|
import json |
|
import os |
|
import re |
|
import shutil |
|
|
|
import torch |
|
from huggingface_hub import snapshot_download |
|
from mergekit.architecture import get_architecture_info |
|
from mergekit.common import ModelReference |
|
from mergekit.io import LazyTensorLoader, TensorWriter |
|
from tqdm import tqdm |
|
|
|
MIXTRAL_MODEL_ID = "mistral-community/Mixtral-8x22B-v0.1" |
|
MIXTRAL_PATH = snapshot_download(repo_id=MIXTRAL_MODEL_ID) |
|
print(f"Mixtral downloaded to: {MIXTRAL_PATH}") |
|
|
|
MISTRAL_PATH = snapshot_download( |
|
repo_id="mistralai/Mistral-7B-v0.1", allow_patterns=["config.json"] |
|
) |
|
print(f"Mistral config downloaded to: {MISTRAL_PATH}") |
|
|
|
with open(os.path.join(MISTRAL_PATH, "config.json"), "r") as f: |
|
mistral_config = json.load(f) |
|
|
|
with open(os.path.join(MIXTRAL_PATH, "config.json"), "r") as f: |
|
mixtral_config = json.load(f) |
|
|
|
combined_config = { |
|
key: mixtral_config[key] for key in mistral_config if key in mixtral_config |
|
} |
|
combined_config["architectures"] = ["MistralForCausalLM"] |
|
combined_config["model_type"] = "mistral" |
|
|
|
mixtral_model_ref = ModelReference.parse(MIXTRAL_PATH) |
|
mixtral_architecture_info = get_architecture_info(mixtral_model_ref.config()) |
|
mixtral_loader = LazyTensorLoader(mixtral_model_ref.tensor_index(), lazy_unpickle=True) |
|
|
|
ALLOW_LIST = ["generation_config.json", "tokenizer.model", "tokenizer_config.json"] |
|
|
|
def copy_directory(src, dest, allowed_patterns): |
|
os.makedirs(dest, exist_ok=True) |
|
for root, dirs, files in os.walk(src): |
|
# Only keep directories that match at least one of the allowed patterns |
|
dirs[:] = [d for d in dirs if any(fnmatch.fnmatch(d, pattern) for pattern in allowed_patterns)] |
|
for file in files: |
|
# Only copy files that match at least one of the allowed patterns |
|
if any(fnmatch.fnmatch(file, pattern) for pattern in allowed_patterns): |
|
src_path = os.path.join(root, file) |
|
dest_path = os.path.join(dest, os.path.relpath(src_path, src)) |
|
os.makedirs(os.path.dirname(dest_path), exist_ok=True) |
|
shutil.copy2(src_path, dest_path) |
|
|
|
def get_tensor(layer_num, expert_num, tensor_type): |
|
weight_name = f"model.layers.{layer_num}.block_sparse_moe.experts.{expert_num}.{tensor_type}.weight" |
|
return mixtral_loader.get_tensor(weight_name) |
|
|
|
|
|
def extract_layer_number(string): |
|
match = re.search(r"layers\.(\d+)\.", string) |
|
return int(match.group(1)) if match else None |
|
|
|
|
|
def save_expert_as_dense(output_path, expert_num): |
|
dense_model_ref = ModelReference.parse(output_path) |
|
dense_architecture_info = get_architecture_info(dense_model_ref.config()) |
|
|
|
writer = TensorWriter(output_path, safe_serialization=True) |
|
|
|
for weight_info in tqdm(dense_architecture_info.all_weights(dense_model_ref.config())): |
|
if weight_info.name.endswith(".up_proj.weight"): |
|
layer_num = extract_layer_number(weight_info.name) |
|
writer.save_tensor(weight_info.name, get_tensor(layer_num, expert_num, "w3")) |
|
elif weight_info.name.endswith(".down_proj.weight"): |
|
layer_num = extract_layer_number(weight_info.name) |
|
writer.save_tensor(weight_info.name, get_tensor(layer_num, expert_num, "w2")) |
|
elif weight_info.name.endswith(".gate_proj.weight"): |
|
layer_num = extract_layer_number(weight_info.name) |
|
writer.save_tensor(weight_info.name, get_tensor(layer_num, expert_num, "w1")) |
|
else: |
|
writer.save_tensor(weight_info.name, mixtral_loader.get_tensor(weight_info.name)) |
|
|
|
writer.finalize() |
|
|
|
|
|
num_experts = mixtral_config["num_local_experts"] |
|
|
|
for expert_num in range(num_experts): |
|
dense_path = f"./dense_expert_{expert_num}" |
|
copy_directory(MIXTRAL_PATH, dense_path, ALLOW_LIST) |
|
|
|
with open(os.path.join(dense_path, "config.json"), "w") as f: |
|
json.dump(combined_config, f, indent=2) |
|
|
|
save_expert_as_dense(dense_path, expert_num) |
|
print(f"Dense model #{expert_num} saved to {os.path.abspath(dense_path)}") |
|
``` |
