MiniCPM implementation (#685)

* Added support for the MiniCPM architecture * Added support for the MiniCPM architecture * Updated utils.py and LORA.md * Updated utils.py and LORA.md * Update implementation details for MiniCPM architecture * Cleaning up * fixed the missing lm.head layer problem * Refactor Model class to dynamically handle tied and untied word embeddings * Quick update * added a dynamic rope scaling base calucaltion * Added support for the MiniCPM architecture * Added support for the MiniCPM architecture * Updated utils.py and LORA.md * Updated utils.py and LORA.md * Update implementation details for MiniCPM architecture * Cleaning up * fixed the missing lm.head layer problem * Refactor Model class to dynamically handle tied and untied word embeddings * added a dynamic rope scaling base calucaltion * quick fix and clean up * clean up again * removed the MiniCPMNorm class as its not used * forgot something, sorry * format * version bump --------- Co-authored-by: Awni Hannun <awni@apple.com>
2025-10-31 02:48:07 +08:00 · 2024-04-26 00:29:28 +02:00
parent 685012c2ad
commit 2c1c9e9024
4 changed files with 251 additions and 22 deletions
--- a/llms/mlx_lm/LORA.md
+++ b/llms/mlx_lm/LORA.md
@@ -11,16 +11,17 @@ LoRA (QLoRA).[^qlora] LoRA fine-tuning works with the following model families:
 - Qwen2
 - Gemma
 - OLMo
+- MiniCPM

 ## Contents

-* [Run](#Run)
-  * [Fine-tune](#Fine-tune)
-  * [Evaluate](#Evaluate)
-  * [Generate](#Generate)
-* [Fuse](#Fuse)
-* [Data](#Data)
-* [Memory Issues](#Memory-Issues)
+- [Run](#Run)
+  - [Fine-tune](#Fine-tune)
+  - [Evaluate](#Evaluate)
+  - [Generate](#Generate)
+- [Fuse](#Fuse)
+- [Data](#Data)
+- [Memory Issues](#Memory-Issues)

 ## Run

@@ -122,7 +123,7 @@ To upload a fused model, supply the `--upload-repo` and `--hf-path` arguments
 to `mlx_lm.fuse`. The latter is the repo name of the original model, which is
 useful for the sake of attribution and model versioning.

-For example, to fuse and upload a model derived from Mistral-7B-v0.1, run: 
+For example, to fuse and upload a model derived from Mistral-7B-v0.1, run:

 ```shell
 mlx_lm.fuse \
@@ -144,38 +145,54 @@ can specify the file name with `--gguf-path`.

 ## Data

-The LoRA command expects you to provide a dataset with `--data`.  The MLX
+The LoRA command expects you to provide a dataset with `--data`. The MLX
 Examples GitHub repo has an [example of the WikiSQL
 data](https://github.com/ml-explore/mlx-examples/tree/main/lora/data) in the
 correct format.

 For fine-tuning (`--train`), the data loader expects a `train.jsonl` and a
 `valid.jsonl` to be in the data directory. For evaluation (`--test`), the data
-loader expects a `test.jsonl` in the data directory. 
+loader expects a `test.jsonl` in the data directory.

 Currently, `*.jsonl` files support three data formats: `chat`,
 `completions`, and `text`. Here are three examples of these formats:

 `chat`:
-  
+
 ```jsonl
-{"messages": [
-  {"role": "system", "content": "You are a helpful assistant." },
-  {"role": "user", "content": "Hello."},
-  {"role": "assistant", "content": "How can I assistant you today."},
-]}
+{
+  "messages": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": "Hello."
+    },
+    {
+      "role": "assistant",
+      "content": "How can I assistant you today."
+    }
+  ]
+}
 ```

 `completions`:
-  
+
 ```jsonl
-{"prompt": "What is the capital of France?", "completion": "Paris."}
+{
+  "prompt": "What is the capital of France?",
+  "completion": "Paris."
+}
 ```

 `text`:

 ```jsonl
-{"text": "This is an example for the model."}
+{
+  "text": "This is an example for the model."
+}
 ```

 Note, the format is automatically determined by the dataset. Note also, keys in
@@ -207,7 +224,7 @@ of memory. Here are some tips to reduce memory use should you need to do so:

 1. Try quantization (QLoRA). You can use QLoRA by generating a quantized model
   with `convert.py` and the `-q` flag. See the [Setup](#setup) section for
-   more details. 
+   more details.

 2. Try using a smaller batch size with `--batch-size`. The default is `4` so
   setting this to `2` or `1` will reduce memory consumption. This may slow
@@ -244,6 +261,5 @@ tokens-per-second, using the MLX Example
 [`wikisql`](https://github.com/ml-explore/mlx-examples/tree/main/lora/data)
 data set.

-
 [^lora]: Refer to the [arXiv paper](https://arxiv.org/abs/2106.09685) for more details on LoRA.
 [^qlora]: Refer to the paper [QLoRA: Efficient Finetuning of Quantized LLMs](https://arxiv.org/abs/2305.14314)
--- a/llms/mlx_lm/models/minicpm.py
+++ b/llms/mlx_lm/models/minicpm.py
@@ -0,0 +1,212 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    dim_model_base: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    max_position_embeddings: int
+    scale_depth: float
+    scale_emb: float
+    rope_theta: float = 1000000.0
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[str, float]]] = None
+    tie_word_embeddings: bool = False
+
+
+class MLP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.gate_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(args.intermediate_size, args.hidden_size, bias=False)
+
+    def __call__(self, x):
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+
+        self.hidden_size = args.hidden_size
+        self.num_heads = n_heads = args.num_attention_heads
+        self.rope_theta = args.rope_theta
+        self.max_position_embeddings = args.max_position_embeddings
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.num_key_value_heads = args.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        rope_scale = (
+            1 / args.rope_scaling["factor"]
+            if args.rope_scaling is not None and args.rope_scaling["type"] == "linear"
+            else 1
+        )
+
+        self.rope = nn.RoPE(
+            dims=self.head_dim,
+            traditional=args.rope_traditional,
+            base=self.rope_theta,
+            scale=rope_scale,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ):
+        B, L, _ = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.num_key_value_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            key_cache, value_cache = cache
+            queries = self.rope(queries, offset=key_cache.shape[2])
+            keys = self.rope(keys, offset=key_cache.shape[2])
+            keys = mx.concatenate([key_cache, keys], axis=2)
+            values = mx.concatenate([value_cache, values], axis=2)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        attn_output = mx.fast.scaled_dot_product_attention(
+            queries, keys, values, scale=self.scale, mask=mask
+        )
+
+        attn_output = attn_output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.o_proj(attn_output), (keys, values)
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.hidden_size = args.hidden_size
+        self.num_hidden_layers = args.num_hidden_layers
+
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+        self.scale_depth = args.scale_depth
+        self.num_hidden_layers = args.num_hidden_layers
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        r, cache = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r * (self.scale_depth / np.sqrt(self.num_hidden_layers))
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r * (self.scale_depth / np.sqrt(self.num_hidden_layers))
+        return out, cache
+
+
+class MiniCPMModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        assert self.vocab_size > 0
+
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [DecoderLayer(args) for _ in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs) * self.args.scale_emb
+
+        mask = None
+        if h.shape[1] > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(h.shape[1])
+            mask = mask.astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for e, layer in enumerate(self.layers):
+            h, cache[e] = layer(h, mask, cache[e])
+
+        return self.norm(h), cache
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = MiniCPMModel(args)
+
+        if not self.args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out, cache = self.model(inputs, cache)
+
+        if not self.args.tie_word_embeddings:
+            out = self.lm_head(out / (self.args.hidden_size / self.args.dim_model_base))
+        else:
+            out = out @ self.model.embed_tokens.weight.T
+
+        return out, cache
+
+    def sanitize(self, weights):
+        if "lm_head.weight" not in weights:
+            weights["lm_head.weight"] = weights["model.embed_tokens.weight"]
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
--- a/llms/mlx_lm/tuner/utils.py
+++ b/llms/mlx_lm/tuner/utils.py
@@ -77,6 +77,7 @@ def linear_to_lora_layers(
        "gemma",
        "starcoder2",
        "cohere",
+        "minicpm",
    ]:
        keys = set(["self_attn.q_proj", "self_attn.v_proj"])
        if model.model_type == "mixtral":
--- a/llms/mlx_lm/version.py
+++ b/llms/mlx_lm/version.py
@@ -1,3 +1,3 @@
 # Copyright © 2023-2024 Apple Inc.

-__version__ = "0.10.0"
+__version__ = "0.12.0"