-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathlowerer.py
More file actions
243 lines (235 loc) · 9.33 KB
/
Copy pathlowerer.py
File metadata and controls
243 lines (235 loc) · 9.33 KB
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
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
"""
Lower HIR -> MIR (very small for MVP).
This will map HIR nodes (Linear/ReLU) to MIR primitive ops (matmul, relu).
"""
from typing import Dict, Any
import uuid
def hir_to_mir(hir_dict: Dict[str, Any]) -> Dict[str, Any]:
# Correct lowering with proper input/output chaining
nodes = []
edges = []
current_tensor = "input" # Start with the input tensor
# We'll emit a node per HIR node
for block in hir_dict.get("blocks", []):
for i, n in enumerate(block.get("nodes", [])):
node_id = n["id"]
op = n["op"]
output_tensor = f"t_{len(nodes)}"
if op == "Linear":
# matmul followed by add (bias) is simplified to 'linear' here (MVP)
nodes.append({
"id": node_id,
"op_type": "linear",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "ReLU":
nodes.append({
"id": node_id,
"op_type": "relu",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": {},
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "Conv2d":
nodes.append({
"id": node_id,
"op_type": "conv2d",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "BatchNorm2d":
nodes.append({
"id": node_id,
"op_type": "batchnorm2d",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "AdaptiveAvgPool2d":
nodes.append({
"id": node_id,
"op_type": "adaptive_avg_pool2d",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "MaxPool2d":
nodes.append({
"id": node_id,
"op_type": "max_pool2d",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "AvgPool2d":
nodes.append({
"id": node_id,
"op_type": "avg_pool2d",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "Flatten":
nodes.append({
"id": node_id,
"op_type": "flatten",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "MultiheadAttention":
# Decompose attention into QKV projection + attention + output projection
params = n.get("params", {})
embed_dim = params.get("embed_dim", 512)
num_heads = params.get("num_heads", 8)
head_dim = embed_dim // num_heads
# QKV projection node
qkv_tensor = f"t_{len(nodes)}_qkv"
nodes.append({
"id": f"{node_id}_qkv_proj",
"op_type": "attention_qkv_projection",
"inputs": [current_tensor],
"outputs": [qkv_tensor],
"params": {
"embed_dim": embed_dim,
"num_heads": num_heads,
"head_dim": head_dim
},
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
# Attention computation node
attn_tensor = f"t_{len(nodes)}_attn"
nodes.append({
"id": f"{node_id}_attention",
"op_type": "scaled_dot_product_attention",
"inputs": [qkv_tensor],
"outputs": [attn_tensor],
"params": {
"num_heads": num_heads,
"head_dim": head_dim,
"dropout": params.get("dropout", 0.0)
},
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
# Output projection node
nodes.append({
"id": f"{node_id}_out_proj",
"op_type": "linear",
"inputs": [attn_tensor],
"outputs": [output_tensor],
"params": {
"in": embed_dim,
"out": embed_dim
},
"compression": {},
"hints": {},
"metadata": {"origin_name": f"{n.get('name')}.out_proj"}
})
elif op == "LayerNorm":
nodes.append({
"id": node_id,
"op_type": "layer_norm",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "GELU":
nodes.append({
"id": node_id,
"op_type": "gelu",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "Embedding":
nodes.append({
"id": node_id,
"op_type": "embedding",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "SiLU":
nodes.append({
"id": node_id,
"op_type": "silu",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": {},
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
elif op == "Mish":
nodes.append({
"id": node_id,
"op_type": "mish",
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": {},
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
else:
nodes.append({
"id": node_id,
"op_type": op.lower(),
"inputs": [current_tensor],
"outputs": [output_tensor],
"params": n.get("params", {}),
"compression": {},
"hints": {},
"metadata": {"origin_name": n.get("name")}
})
# Update current_tensor for the next node
current_tensor = output_tensor
mir = {
"version": "itc-mir-v0.1",
"graph": {
"nodes": nodes,
"edges": [] # edges are implicit in MVP (can be added later)
},
"metadata": {"origin": f"{hir_dict.get('name')}"}
}
return mir