Upload folder using huggingface_hub

README.md

@@ -0,0 +1,28 @@
+---
+language:
+- fr
+- it
+- de
+- es
+- en
+license: apache-2.0
+tags:
+- moe
+- mlx
+---
+
+# mlx-community/mixtral-8x22b-4bit
+This model was converted to MLX format from [`v2ray/Mixtral-8x22B-v0.1`]() using mlx-lm version **0.4.0**.
+Refer to the [original model card](https://huggingface.co/v2ray/Mixtral-8x22B-v0.1) for more details on the model.
+## Use with mlx
+
+```bash
+pip install mlx-lm
+```
+
+```python
+from mlx_lm import load, generate
+
+model, tokenizer = load("mlx-community/mixtral-8x22b-4bit")
+response = generate(model, tokenizer, prompt="hello", verbose=True)
+```

config.json

@@ -0,0 +1,88 @@
+{
+    "add_cross_attention": false,
+    "architectures": [
+        "MixtralForCausalLM"
+    ],
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 1,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_act": "silu",
+    "hidden_size": 6144,
+    "id2label": {
+        "0": "LABEL_0",
+        "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 16384,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+        "LABEL_0": 0,
+        "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 65536,
+    "min_length": 0,
+    "model_type": "mixtral",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 48,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_experts_per_tok": 2,
+    "num_hidden_layers": 56,
+    "num_key_value_heads": 8,
+    "num_local_experts": 8,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_router_logits": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "quantization": {
+        "group_size": 64,
+        "bits": 4
+    },
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rms_norm_eps": 1e-05,
+    "rope_theta": 1000000,
+    "router_aux_loss_coef": 0.001,
+    "router_jitter_noise": 0.0,
+    "sep_token_id": null,
+    "sliding_window": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "bfloat16",
+    "torchscript": false,
+    "transformers_version": "4.39.3",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 32000
+}

convert.py

@@ -0,0 +1,278 @@
+# Copyright 2023 Mistral AI and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import json
+import os
+
+import torch
+from safetensors.torch import load_file
+
+from transformers import (
+    MixtralConfig,
+    MixtralForCausalLM,
+)
+
+"""
+Sample usage:
+
+```
+python src/transformers/models/mixtral/convert_mixtral_weights_to_hf.py \
+    --input_dir /path/to/downloaded/mixtral/weights --model_size 7B --output_dir /output/path
+```
+
+Thereafter, models can be loaded via:
+
+```py
+from transformers import MixtralForCausalLM
+
+model = MixtralForCausalLM.from_pretrained("/output/path")
+```
+
+Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
+come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
+"""
+
+def compute_intermediate_size(n, ffn_dim_multiplier=1, multiple_of=256):
+    return multiple_of * ((int(ffn_dim_multiplier * int(8 * n / 3)) + multiple_of - 1) // multiple_of)
+
+def read_json(path):
+    with open(path, "r") as f:
+        return json.load(f)
+
+def write_json(text, path):
+    with open(path, "w") as f:
+        json.dump(text, f)
+
+def write_model(model_path, input_base_path, model_size, safe_serialization=True):
+    os.makedirs(model_path, exist_ok=True)
+
+    params = read_json(os.path.join(input_base_path, "params.json"))
+    num_shards = 1
+
+    # For some reason this is a string in the params.json
+    sliding_window = int(params["sliding_window"]) if "sliding_window" in params else None
+    base = params.get("rope_theta", 10000.0)
+    vocab_size = params["vocab_size"]
+
+    if model_size == "7B":
+        dim = params["hidden_size"]
+        max_position_embeddings = 4096 * 8
+        num_local_experts = params["num_local_experts"]
+        ffn_dim = params["intermediate_size"]
+        n_layers = params["num_hidden_layers"]
+        n_heads = params["num_attention_heads"]
+        n_heads_per_shard = n_heads // num_shards
+        dims_per_head = dim // n_heads
+        if "num_key_value_heads" in params:
+            num_key_value_heads = params["num_key_value_heads"] # for GQA / MQA
+            num_local_key_value_heads = num_key_value_heads // num_shards
+            key_value_dim = dims_per_head * num_local_key_value_heads
+        else: # compatibility with other checkpoints
+            num_key_value_heads = n_heads
+            num_local_key_value_heads = n_heads_per_shard
+            key_value_dim = dim
+        rms_norm_eps = params["rms_norm_eps"]
+    elif model_size == "22B":
+        dim = params["dim"]
+        max_position_embeddings = params["max_seq_len"]
+        num_local_experts = params["moe"]["num_experts"]
+        ffn_dim = params["hidden_dim"]
+        n_layers = params["n_layers"]
+        n_heads = params["n_heads"]
+        n_heads_per_shard = n_heads // num_shards
+        dims_per_head = dim // n_heads
+        if "n_kv_heads" in params:
+            num_key_value_heads = params["n_kv_heads"] # for GQA / MQA
+            num_local_key_value_heads = num_key_value_heads // num_shards
+            key_value_dim = dims_per_head * num_local_key_value_heads
+        else: # compatibility with other checkpoints
+            num_key_value_heads = n_heads
+            num_local_key_value_heads = n_heads_per_shard
+            key_value_dim = dim
+        rms_norm_eps = params["norm_eps"]
+    else:
+        raise Exception("Illegal model size:", model_size)
+
+    # permute for sliced rotary
+    def permute(w, n_heads=n_heads, dim1=dim, dim2=dim):
+        return w.view(n_heads, dim1 // n_heads // 2, 2, dim2).transpose(1, 2).reshape(dim1, dim2)
+
+    print(f"Fetching all parameters from the checkpoint at \"{input_base_path}\"...")
+    # Load weights
+    if model_size == "7B":
+        loaded = [
+            torch.load(os.path.join(input_base_path, f"consolidated.{i:02d}.pt"), map_location="cpu") for i in range(8)
+        ]
+        merged_state_dict = {}
+        for state_dict in loaded:
+            merged_state_dict.update(state_dict)
+    elif model_size == "22B":
+        merged_state_dict = load_file(os.path.join(input_base_path, "consolidated.safetensors"))
+    print("Parameters load finished.")
+
+    state_dict = {}
+    for layer_i in range(n_layers):
+        print(f"At layer {layer_i}...")
+        # Sharded
+        # Note that attention.w{q,k,v,o}, feed_fordward.w[1,2,3], attention_norm.weight and ffn_norm.weight share
+        # the same storage object, saving attention_norm and ffn_norm will save other weights too, which is
+        # redundant as other weights will be stitched from multiple shards. To avoid that, they are cloned.
+
+        state_dict.update(
+            {
+                f"model.layers.{layer_i}.input_layernorm.weight": merged_state_dict[
+                    f"layers.{layer_i}.attention_norm.weight"
+                ].clone(),
+                f"model.layers.{layer_i}.post_attention_layernorm.weight": merged_state_dict[
+                    f"layers.{layer_i}.ffn_norm.weight"
+                ].clone(),
+            }
+        )
+
+        state_dict[f"model.layers.{layer_i}.self_attn.q_proj.weight"] = permute(
+            merged_state_dict[f"layers.{layer_i}.attention.wq.weight"]
+            .view(n_heads_per_shard, dims_per_head, dim)
+            .reshape(dim, dim)
+        )
+        state_dict[f"model.layers.{layer_i}.self_attn.k_proj.weight"] = permute(
+            merged_state_dict[f"layers.{layer_i}.attention.wk.weight"]
+            .view(num_local_key_value_heads, dims_per_head, dim)
+            .reshape(key_value_dim, dim),
+            num_key_value_heads,
+            key_value_dim,
+            dim,
+        )
+        state_dict[f"model.layers.{layer_i}.self_attn.v_proj.weight"] = (
+            merged_state_dict[f"layers.{layer_i}.attention.wv.weight"]
+            .view(num_local_key_value_heads, dims_per_head, dim)
+            .reshape(key_value_dim, dim)
+        )
+
+        state_dict[f"model.layers.{layer_i}.self_attn.o_proj.weight"] = merged_state_dict[
+            f"layers.{layer_i}.attention.wo.weight"
+        ]
+
+        if model_size == "7B":
+            w1 = merged_state_dict[f"layers.{layer_i}.block_sparse_moe.w1"]
+            w2 = merged_state_dict[f"layers.{layer_i}.block_sparse_moe.w2"]
+            w3 = merged_state_dict[f"layers.{layer_i}.block_sparse_moe.w3"]
+
+            experts_w1 = [
+                w1[ffn_dim * expert_idx : ffn_dim * (expert_idx + 1), :].contiguous().clone()
+                for expert_idx in range(num_local_experts)
+            ]
+
+            for idx, expert_block in enumerate(experts_w1):
+                expert_key = f"model.layers.{layer_i}.block_sparse_moe.experts.{idx}.w1"
+                state_dict[expert_key + ".weight"] = expert_block.clone()
+
+            experts_w2 = [
+                w2[ffn_dim * expert_idx : ffn_dim * (expert_idx + 1), :].contiguous().clone()
+                for expert_idx in range(num_local_experts)
+            ]
+
+            for idx, expert_block in enumerate(experts_w2):
+                expert_key = f"model.layers.{layer_i}.block_sparse_moe.experts.{idx}.w2"
+                state_dict[expert_key + ".weight"] = expert_block.T.clone().contiguous()
+
+            experts_w3 = [
+                w3[ffn_dim * expert_idx : ffn_dim * (expert_idx + 1), :].contiguous().clone()
+                for expert_idx in range(num_local_experts)
+            ]
+
+            for idx, expert_block in enumerate(experts_w3):
+                expert_key = f"model.layers.{layer_i}.block_sparse_moe.experts.{idx}.w3"
+                state_dict[expert_key + ".weight"] = expert_block.clone()
+
+            state_dict[f"model.layers.{layer_i}.block_sparse_moe.gate.weight"] = merged_state_dict[
+                f"layers.{layer_i}.block_sparse_moe.gate.weight"
+            ]
+        elif model_size == "22B":
+            for expert_i in range(num_local_experts):
+                w1 = merged_state_dict[f"layers.{layer_i}.feed_forward.experts.{expert_i}.w1.weight"]
+                w2 = merged_state_dict[f"layers.{layer_i}.feed_forward.experts.{expert_i}.w2.weight"]
+                w3 = merged_state_dict[f"layers.{layer_i}.feed_forward.experts.{expert_i}.w3.weight"]
+                state_dict[f"model.layers.{layer_i}.block_sparse_moe.experts.{expert_i}.w1.weight"] = w1.contiguous().clone()
+                state_dict[f"model.layers.{layer_i}.block_sparse_moe.experts.{expert_i}.w2.weight"] = w2.contiguous().clone()
+                state_dict[f"model.layers.{layer_i}.block_sparse_moe.experts.{expert_i}.w3.weight"] = w3.contiguous().clone()
+            state_dict[f"model.layers.{layer_i}.block_sparse_moe.gate.weight"] = merged_state_dict[
+                f"layers.{layer_i}.feed_forward.gate.weight"
+            ]
+
+    state_dict.update(
+        {
+            "model.norm.weight": merged_state_dict["norm.weight"],
+            "model.embed_tokens.weight": merged_state_dict["tok_embeddings.weight"],
+            "lm_head.weight": merged_state_dict["output.weight"],
+        }
+    )
+
+    config_additional_kwargs = {}
+    if model_size == "22B":
+        config_additional_kwargs["num_experts_per_tok"] = params["moe"]["num_experts_per_tok"]
+    config = MixtralConfig(
+        hidden_size=dim,
+        intermediate_size=ffn_dim,
+        num_attention_heads=n_heads,
+        num_hidden_layers=n_layers,
+        rms_norm_eps=rms_norm_eps,
+        num_key_value_heads=num_key_value_heads,
+        vocab_size=vocab_size,
+        rope_theta=base,
+        max_position_embeddings=max_position_embeddings,
+        sliding_window=sliding_window,
+        num_local_experts=num_local_experts,
+        **config_additional_kwargs
+    )
+
+    print("Loading the checkpoint in a Mixtral model.")
+    with torch.device("meta"):
+        model = MixtralForCausalLM(config)
+    # Avoid saving this as part of the config.
+    del model.config._name_or_path
+    model.config.torch_dtype = torch.bfloat16
+    print("Saving in the Transformers format.")
+
+    model.load_state_dict(state_dict, strict=True, assign=True)
+
+    for n, p in model.named_parameters():
+        assert p.device.type != "meta", f"{n} has not been loaded!"
+
+    model.save_pretrained(model_path, safe_serialization=safe_serialization)
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input-dir",
+        help="Location of Mixtral weights, which contains tokenizer.model and model folders",
+        required=True,
+    )
+    parser.add_argument(
+        "--model-size",
+        choices=["7B", "22B"],
+        help="'f' models correspond to the finetuned versions, and are specific to the Mixtral official release. For more details on Mixtral, checkout the original repo: https://huggingface.co/mistral-ai",
+        default="7B",
+    )
+    parser.add_argument("--output-dir", help="Location to write HF model", required=True)
+    parser.add_argument("--safe-serialization", type=bool, default=True, help="Whether or not to save using `safetensors`.")
+    args = parser.parse_args()
+    write_model(
+        model_path=args.output_dir,
+        input_base_path=args.input_dir,
+        model_size=args.model_size,
+        safe_serialization=args.safe_serialization,
+    )
+
+if __name__ == "__main__":
+    main()

model-00001-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:314b7c3ef0ac39ffa20bfc29fcc3ffb355fcb9a0dbc456b8d86f075bb03251ae
+size 5348243527

model-00002-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:387e2cd7fe5855521343788e960a2a0ec45b1f0a48fe0d0cf40f2920b2e184a1
+size 5351416675

model-00003-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93256e72ed9577dd863619d12ab12b44d7f802964f993b742b4f83ea8c7edfd8
+size 5351416739

model-00004-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2925dc99a4bb8a098f12190968d8dbfa6b2992f6c3b5d2ee92039925cf39c4c
+size 5358370680

model-00005-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdc080ecf96f9ecd22f3d723278a4d33a46c176f6a5b1dbe276929651701beb2
+size 5351416988

model-00006-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51c96a3b9cea6c47e37736bc12809e1d36870ca137c907100b2c95244fe9c872
+size 5351417004

model-00007-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11f578df7d003f4a32187d6ab3e3c50a2edf82eb9542ec37a233a0af8d614256
+size 5351416986

model-00008-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecc9bc38ac8b2cdde50491c0982861a5ddcff5679816c43aa1e8fdb25c88c34d
+size 5351416946

model-00009-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9619c42f210da268126cb9aef2d529d7546005a18864f3b549650167061930c2
+size 5358370676

model-00010-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f809b7cac05afcd9050f26ab38e77eed2a6c9e56af0ef4dea9db8977f03e994c
+size 5351417016

model-00011-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e89876f56dc63477ddba26c708b0c100ab42c17c27de9abc6b9f8061283d25ff
+size 5351417002

model-00012-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:087f926e66d6ec285f4570c00d61507efd9db949e6738ce4eb755f90099a9415
+size 5351416958

model-00013-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f06e87f350b4da91fa163ac1429096a76b4808b0a7c46748fa5d14e957e090f0
+size 5351416996

model-00014-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:588f928d1787312f0f532e6825392d96fb5f2735aca3e3eb2cd8dafc7d3eaf08
+size 5358370654

model-00015-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0032391c95d00dd74b464e9b66ab3deb489cac718d1db09e0481674d9f127a5
+size 4449777275

special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [],
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": null,
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}