[object Object]

← back to Exo

add support for qwen2.5, initially adding mlx-community/Qwen2.5-14B-Instruct-4bit

3597fba35638627e2d60ce05113b8ce1003282ef · 2024-09-18 22:49:24 +0100 · Alex Cheema

Files touched

Diff

commit 3597fba35638627e2d60ce05113b8ce1003282ef
Author: Alex Cheema <alexcheema123@gmail.com>
Date:   Wed Sep 18 22:49:24 2024 +0100

    add support for qwen2.5, initially adding mlx-community/Qwen2.5-14B-Instruct-4bit
---
 exo/inference/mlx/models/qwen2.py     | 127 ++++++++++++++++++++++++++++++++++
 exo/models.py                         |   4 ++
 tinychat/examples/tinychat/index.html |   1 +
 3 files changed, 132 insertions(+)

diff --git a/exo/inference/mlx/models/qwen2.py b/exo/inference/mlx/models/qwen2.py
new file mode 100644
index 00000000..7aed2d04
--- /dev/null
+++ b/exo/inference/mlx/models/qwen2.py
@@ -0,0 +1,127 @@
+from dataclasses import dataclass, field
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from mlx_lm.models.base import create_attention_mask
+from mlx_lm.models.qwen2 import TransformerBlock, ModelArgs
+
+from ...shard import Shard
+from .base import IdentityBlock
+
+
+@dataclass
+class ModelArgs(ModelArgs):
+  shard: Shard = field(default_factory=lambda: Shard("", 0, 0, 0))
+
+  def __post_init__(self):
+    super().__post_init__()  # Ensure parent initializations are respected
+
+    if isinstance(self.shard, Shard):
+      return
+    if not isinstance(self.shard, dict):
+      raise TypeError(f"Expected shard to be a Shard instance or a dict, got {type(self.shard)} instead")
+
+    self.shard = Shard(**self.shard)
+
+class Qwen2Model(nn.Module):
+  def __init__(self, args: ModelArgs):
+    super().__init__()
+    self.args = args
+    self.vocab_size = args.vocab_size
+    self.num_hidden_layers = args.num_hidden_layers
+    assert self.vocab_size > 0
+    if self.args.shard.is_first_layer():
+      self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+    self.layers = []
+    for i in range(self.num_hidden_layers):
+      if self.args.shard.start_layer <= i <= self.args.shard.end_layer:
+        self.layers.append(TransformerBlock(args=args))
+      else:
+        self.layers.append(IdentityBlock())
+    if self.args.shard.is_last_layer():
+      self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+  def __call__(
+    self,
+    inputs: mx.array,
+    cache=None,
+  ):
+    if self.args.shard.is_first_layer():
+      h = self.embed_tokens(inputs)
+    else:
+      h = inputs
+
+    mask = None
+    if h.shape[1] > 1:
+      mask = create_attention_mask(h, cache)
+
+    if cache is None:
+      cache = [None] * len(self.layers)
+
+    for layer, c in zip(self.layers, cache):
+      h = layer(h, mask, c)
+
+    if self.args.shard.is_last_layer():
+      h = self.norm(h)
+    return h
+
+
+class Model(nn.Module):
+  def __init__(self, args: ModelArgs):
+    super().__init__()
+    self.args = args
+    self.model_type = args.model_type
+    self.model = Qwen2Model(args)
+    if self.args.shard.is_last_layer():
+      if not 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 = self.model(inputs, cache)
+    if self.args.shard.is_last_layer():
+      if self.args.tie_word_embeddings:
+        out = self.model.embed_tokens.as_linear(out)
+      else:
+        out = self.lm_head(out)
+    return out
+
+  def sanitize(self, weights):
+    shard_state_dict = {}
+
+    for key, value in weights.items():
+      if "self_attn.rotary_emb.inv_freq" in key:
+        continue
+      if key.startswith('model.layers.'):
+        layer_num = int(key.split('.')[2])
+        if self.args.shard.start_layer <= layer_num <= self.args.shard.end_layer:
+          shard_state_dict[key] = value
+      elif self.args.shard.is_first_layer() and key.startswith('model.embed_tokens'):
+        shard_state_dict[key] = value
+      elif (self.args.shard.is_last_layer() and self.args.tie_word_embeddings) and key.startswith('model.embed_tokens'):
+        shard_state_dict[key] = value
+      elif (self.args.shard.is_last_layer() and not self.args.tie_word_embeddings) and key.startswith('lm_head'):
+        shard_state_dict[key] = value
+      elif self.args.shard.is_last_layer() and (key.startswith('model.norm')):
+        shard_state_dict[key] = value
+
+    if self.args.tie_word_embeddings:
+      shard_state_dict.pop("lm_head.weight", None)
+
+    return shard_state_dict
+
+  @property
+  def layers(self):
+    return self.model.layers
+
+  @property
+  def head_dim(self):
+    return self.args.hidden_size // self.args.num_attention_heads
+
+  @property
+  def n_kv_heads(self):
+    return self.args.num_key_value_heads
diff --git a/exo/models.py b/exo/models.py
index 605689b2..3fe95d6b 100644
--- a/exo/models.py
+++ b/exo/models.py
@@ -30,4 +30,8 @@ model_base_shards = {
   "deepseek-coder-v2-lite": {"MLXDynamicShardInferenceEngine": Shard(model_id="mlx-community/DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx", start_layer=0, end_layer=0, n_layers=27),},
   ### llava
   "llava-1.5-7b-hf": {"MLXDynamicShardInferenceEngine": Shard(model_id="llava-hf/llava-1.5-7b-hf", start_layer=0, end_layer=0, n_layers=32),},
+  ### qwen
+  "qwen-2.5-14b": {
+    "MLXDynamicShardInferenceEngine": Shard(model_id="mlx-community/Qwen2.5-14B-Instruct-4bit", start_layer=0, end_layer=0, n_layers=48),
+  },
 }
diff --git a/tinychat/examples/tinychat/index.html b/tinychat/examples/tinychat/index.html
index 59be462c..4a7a95f3 100644
--- a/tinychat/examples/tinychat/index.html
+++ b/tinychat/examples/tinychat/index.html
@@ -37,6 +37,7 @@
 <option value="mistral-large">Mistral Large</option>
 <option value="deepseek-coder-v2-lite">Deepseek Coder V2 Lite</option>
 <option value="llava-1.5-7b-hf">LLaVa 1.5 7B (Vision Model)</option>
+<option value="qwen-2.5-14b">Qwen 2.5 14B</option>
 </select>
 </div>
 <div @popstate.window="

← b39a251d fix: remove extraneous '/'  ·  back to Exo  ·  add more qwen2.5 models: mlx-community/Qwen2.5-7B-Instruct-4 dee83e48 →