File size: 8,089 Bytes
0b0ec56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
# gemma4_optimization.py

"""
Hayson Cheung, 2026, Oringinal Script written to optimize
Gemma4 on Hugging Face's Transformers library. 

LICENSED UNDER THE MIT LICENSE.

This file contains optimized variants of Gemma4 text model components, including a mixin for remapping weights from original Gemma4 models to optimized versions. The optimizations include support for an additional zero-compute expert in the MoE router and experts, as well as adjustments to the router's projection and scaling parameters to accommodate the expanded expert set. The load_optimization_weights method enables loading weights from a base Gemma4 model while remapping tensors as needed for the optimized architecture.
"""

from __future__ import annotations

from dataclasses import dataclass

import torch
from torch import nn

from .modeling_gemma4 import (
    Gemma4ForCausalLM,
    Gemma4TextDecoderLayer,
    Gemma4TextExperts,
    Gemma4TextModel,
    Gemma4TextRouter,
)


@dataclass(frozen=True)
class Gemma4OptimizationLoadResult:
    loaded_keys: tuple[str, ...]
    skipped_keys: tuple[str, ...]

    @property
    def loaded_count(self) -> int:
        return len(self.loaded_keys)

    @property
    def skipped_count(self) -> int:
        return len(self.skipped_keys)


class Gemma4OptimizationWeightsMixin:
    """
    Mixin for modules that need a custom remount step when loading weights
    from an original Gemma4 model into an optimized variant.
    """

    def _remap_optimization_tensors(
        self,
        base_state_dict: dict[str, torch.Tensor],
        target_state_dict: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
        return {}

    def load_optimization_weights(self, base_model: nn.Module) -> Gemma4OptimizationLoadResult:
        if not isinstance(self, nn.Module):
            raise TypeError("Gemma4OptimizationWeightsMixin can only be used with nn.Module subclasses.")
        if not isinstance(base_model, nn.Module):
            raise TypeError("base_model must be an nn.Module.")

        target_state_dict = self.state_dict()
        loaded: dict[str, torch.Tensor] = {}

        for module_name, module in self.named_modules():
            if not isinstance(module, Gemma4OptimizationWeightsMixin):
                continue

            try:
                base_module = base_model if module_name == "" else base_model.get_submodule(module_name)
            except AttributeError:
                continue

            remapped_tensors = module._remap_optimization_tensors(base_module.state_dict(), module.state_dict())
            for tensor_name, tensor_value in remapped_tensors.items():
                full_name = f"{module_name}.{tensor_name}" if module_name else tensor_name
                loaded[full_name] = tensor_value.to(
                    device=target_state_dict[full_name].device,
                    dtype=target_state_dict[full_name].dtype,
                )

        for tensor_name, tensor_value in base_model.state_dict().items():
            if tensor_name in loaded:
                continue
            target_tensor = target_state_dict.get(tensor_name)
            if target_tensor is None or target_tensor.shape != tensor_value.shape:
                continue
            loaded[tensor_name] = tensor_value.to(device=target_tensor.device, dtype=target_tensor.dtype)

        self.load_state_dict(loaded, strict=False)

        skipped = tuple(sorted(set(base_model.state_dict()) - set(loaded)))
        return Gemma4OptimizationLoadResult(tuple(sorted(loaded)), skipped)

    def _load_weights(self, base_model: nn.Module) -> Gemma4OptimizationLoadResult:
        return self.load_optimization_weights(base_model)


def get_total_optimized_experts(num_experts: int, add_zero_compute_expert: bool) -> int:
    return num_experts + int(add_zero_compute_expert)


class OptimizedGemma4TextExperts(Gemma4TextExperts):
    def __init__(self, config):
        super().__init__(config)
        self.total_num_experts = get_total_optimized_experts(
            self.num_experts, getattr(config, "add_zero_compute_expert", False)
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        top_k_index: torch.Tensor,
        top_k_weights: torch.Tensor,
    ) -> torch.Tensor:
        final_hidden_states = torch.zeros_like(hidden_states)
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.total_num_experts)
            expert_mask = expert_mask.permute(2, 1, 0)
            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()

        for expert_idx in expert_hit:
            expert_idx = expert_idx[0]
            if expert_idx >= self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
            current_state = hidden_states[token_idx]
            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
            current_hidden_states = self.act_fn(gate) * up
            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))

        return final_hidden_states


class OptimizedGemma4TextRouter(Gemma4OptimizationWeightsMixin, Gemma4TextRouter):
    def __init__(self, config):
        super().__init__(config)
        self.num_experts = config.num_experts
        self.total_num_experts = get_total_optimized_experts(
            self.num_experts, getattr(config, "add_zero_compute_expert", False)
        )
        self.proj = nn.Linear(config.hidden_size, self.total_num_experts, bias=False)
        self.per_expert_scale = nn.Parameter(torch.ones(self.total_num_experts))

    def _remap_optimization_tensors(
        self,
        base_state_dict: dict[str, torch.Tensor],
        target_state_dict: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
        remapped: dict[str, torch.Tensor] = {}

        base_proj = base_state_dict.get("proj.weight")
        target_proj = target_state_dict.get("proj.weight")
        if (
            base_proj is not None
            and target_proj is not None
            and target_proj.shape[1] == base_proj.shape[1]
            and target_proj.shape[0] == base_proj.shape[0] + 1
        ):
            expanded_proj = target_proj.clone()
            expanded_proj.zero_()
            expanded_proj[: base_proj.shape[0]].copy_(base_proj)
            remapped["proj.weight"] = expanded_proj

        base_per_expert_scale = base_state_dict.get("per_expert_scale")
        target_per_expert_scale = target_state_dict.get("per_expert_scale")
        if (
            base_per_expert_scale is not None
            and target_per_expert_scale is not None
            and target_per_expert_scale.shape[0] == base_per_expert_scale.shape[0] + 1
        ):
            expanded_per_expert_scale = target_per_expert_scale.clone()
            expanded_per_expert_scale.fill_(1.0)
            expanded_per_expert_scale[: base_per_expert_scale.shape[0]].copy_(base_per_expert_scale)
            remapped["per_expert_scale"] = expanded_per_expert_scale

        return remapped


class OptimizedGemma4TextDecoderLayer(Gemma4TextDecoderLayer):
    router_class = OptimizedGemma4TextRouter
    experts_class = OptimizedGemma4TextExperts


class OptimizedGemma4TextModel(Gemma4OptimizationWeightsMixin, Gemma4TextModel):
    decoder_layer_class = OptimizedGemma4TextDecoderLayer


class OptimizedGemma4ForCausalLM(Gemma4OptimizationWeightsMixin, Gemma4ForCausalLM):
    text_model_class = OptimizedGemma4TextModel


__all__ = [
    "Gemma4OptimizationLoadResult",
    "Gemma4OptimizationWeightsMixin",
    "OptimizedGemma4ForCausalLM",
    "OptimizedGemma4TextDecoderLayer",
    "OptimizedGemma4TextExperts",
    "OptimizedGemma4TextModel",
    "OptimizedGemma4TextRouter",
    "get_total_optimized_experts",
]