Create README.md

README.md

@@ -0,0 +1,109 @@
+# Upcycled-Qwen1.5-MoE2.7B
+
+This an attemp (probably too naive) to reproduce the upcycling process used to initialize [Qwen1.5-MoE-A2.7B](https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B) using [Qwen1.5-1.8B](Qwen/Qwen1.5-1.8B).
+
+## Upcycling script
+
+```python
+from torch import nn
+from transformers import AutoModelForCausalLM
+from dataclasses import dataclass
+from transformers import AutoModel
+from typing_extensions import Self
+from copy import deepcopy
+
+@dataclass
+class UpcyclingConfig:
+    finegrained_experts: int
+    partitions_from_mlp: int
+
+    @property
+    def upcycling_factor(self) -> int:
+        return self.finegrained_experts // self.partitions_from_mlp
+
+
+def iterate_in_chunks(list1, list2, chunk_size1, chunk_size2):
+    iterations = max(len(list1) // chunk_size1, len(list2) // chunk_size2)
+    for i in range(iterations):
+        start_idx1 = i * chunk_size1
+        end_idx1 = start_idx1 + chunk_size1
+        start_idx2 = i * chunk_size2
+        end_idx2 = start_idx2 + chunk_size2
+        yield (list1[start_idx1:end_idx1], list2[start_idx2:end_idx2])
+
+
+def chunk_linear(linear: nn.Linear, chunks: int, down_proj: bool = False) -> tuple[nn.Linear, ...]:
+    if not down_proj:
+        in_features = linear.in_features
+        out_features = linear.out_features // chunks
+    else:
+        in_features = linear.in_features // chunks
+        out_features = linear.out_features
+
+    weights = linear.weight.chunk(chunks)
+    biases = linear.bias.chunk(chunks) if linear.bias is not None else [None] * chunks
+    linear_layers = []
+    for weight, bias in zip(weights, biases):
+        new_linear = nn.Linear(
+            in_features=in_features, out_features=out_features, bias=bias is not None
+        )
+        new_linear.weight = nn.Parameter(weight.clone())  # Clone weights to ensure they are not shared
+        if bias is not None:
+            new_linear.bias = nn.Parameter(bias.clone())  # Clone bias if it exists
+        linear_layers.append(new_linear)
+    return tuple(linear_layers)
+
+
+class UpcycledModelMixin:
+    sparse_moe_block_cls: type
+
+    @classmethod
+    def upcycled_from(cls, source_model, config: UpcyclingConfig) -> Self:
+        upcycled_model_config = cls.config_class(**source_model.config.to_dict())
+        if hasattr(upcycled_model_config, "shared_expert_intermediate_size"):
+            upcycled_model_config.shared_expert_intermediate_size = source_model.config.intermediate_size
+
+        upcycled_model = cls(upcycled_model_config)
+        upcycled_model.model.embed_tokens = source_model.model.embed_tokens
+
+        for upcycled_layer, layer in zip(upcycled_model.model.layers, source_model.model.layers):
+            upcycled_layer.self_attn = layer.self_attn
+            upcycled_mlp_layers = [deepcopy(layer.mlp) for _ in range(config.upcycling_factor)]
+
+            if hasattr(upcycled_layer.mlp, "shared_expert"):
+                upcycled_layer.mlp.shared_expert = upcycled_mlp_layers.pop(-1)
+
+            for experts, mlp in iterate_in_chunks(upcycled_layer.mlp.experts, upcycled_mlp_layers, 4, 1):
+                gate_projs = chunk_linear(mlp[0].gate_proj, 4, down_proj=False)
+                up_projs = chunk_linear(mlp[0].up_proj, 4, down_proj=False)
+                down_projs = chunk_linear(mlp[0].down_proj, 4, down_proj=True)
+                for i, expert in enumerate(experts):
+                    expert.gate_proj = gate_projs[i]
+                    expert.up_proj = up_projs[i]
+                    expert.down_proj = down_projs[i]
+                    expert.act_fn = deepcopy(mlp[0].act_fn)
+
+            upcycled_layer.input_layernorm = layer.input_layernorm
+            upcycled_layer.post_attention_layernorm = layer.post_attention_layernorm
+
+        upcycled_model.lm_head = source_model.lm_head
+        return upcycled_model
+
+
+from transformers import Qwen2MoeForCausalLM as _Qwen2MoeForCausalLM
+from transformers.models.qwen2.modeling_qwen2 import Qwen2MLP
+from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+class Qwen2MoeForCausalLM(UpcycledModelMixin, _Qwen2MoeForCausalLM):
+    sparse_moe_block_cls = Qwen2MoeSparseMoeBlock
+
+
+source_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-1.8B")
+model = Qwen2MoeForCausalLM.upcycled_from(
+    source_model,
+    UpcyclingConfig(
+        finegrained_experts=64,
+        partitions_from_mlp=4,
+    ),
+)
+```