File size: 8,591 Bytes
7d52396
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import pytest
import sys


def copy_mlp(llama_mlp, orig_llama_mlp) -> None:
    orig_llama_mlp.w1.weight.copy_(llama_mlp.c_fc1.weight)
    orig_llama_mlp.w3.weight.copy_(llama_mlp.c_fc2.weight)
    orig_llama_mlp.w2.weight.copy_(llama_mlp.c_proj.weight)


def copy_attention(llama_attn, orig_llama_attn) -> None:
    n_embd = llama_attn.c_attn.weight.shape[1]
    orig_llama_attn.wq.weight.copy_(llama_attn.c_attn.weight[:n_embd])
    orig_llama_attn.wk.weight.copy_(llama_attn.c_attn.weight[n_embd:-n_embd])
    orig_llama_attn.wv.weight.copy_(llama_attn.c_attn.weight[-n_embd:])
    orig_llama_attn.wo.weight.copy_(llama_attn.c_proj.weight)


def copy_block(llama_block, orig_llama_block) -> None:
    orig_llama_block.attention_norm.weight.copy_(llama_block.rms_1.scale)
    copy_attention(llama_block.attn, orig_llama_block.attention)
    orig_llama_block.ffn_norm.weight.copy_(llama_block.rms_2.scale)
    copy_mlp(llama_block.mlp, orig_llama_block.feed_forward)


def copy_weights(llama_model, orig_llama_model) -> None:
    orig_llama_model.tok_embeddings.weight.copy_(llama_model.transformer.wte.weight)
    for llama_block, orig_llama_block in zip(llama_model.transformer.h, orig_llama_model.layers):
        copy_block(llama_block, orig_llama_block)
    orig_llama_model.norm.weight.copy_(llama_model.transformer.ln_f.scale)
    orig_llama_model.output.weight.copy_(llama_model.lm_head.weight)


@torch.no_grad()
@pytest.mark.parametrize("kv_cache", (False, True))
def test_to_orig_llama(lit_llama, orig_llama, kv_cache) -> None:
    block_size = 64
    vocab_size = 32000
    n_layer = 16
    n_head = 16
    n_embd = 32
    batch_size = 3

    llama_config = lit_llama.LLaMAConfig(
        block_size=block_size, vocab_size=vocab_size, n_layer=n_layer, n_head=n_head, n_embd=n_embd
    )
    orig_llama_config = orig_llama.ModelArgs(
        dim=n_embd,
        n_layers=n_layer,
        n_heads=n_head,
        vocab_size=vocab_size,
        norm_eps=1e-5,
        max_seq_len=block_size,
        max_batch_size=batch_size,
    )

    seq_len = orig_llama_config.max_seq_len
    token_sample = torch.randint(0, orig_llama_config.vocab_size, size=(batch_size, seq_len), dtype=torch.int64)

    llama_model = lit_llama.LLaMA(llama_config)
    llama_model.apply(llama_model._init_weights)
    orig_llama_model = orig_llama.Transformer(orig_llama_config)

    copy_weights(llama_model, orig_llama_model)

    orig_llama_embed = orig_llama_model.tok_embeddings(token_sample)
    llama_embed = llama_model.transformer.wte(token_sample)
    assert torch.allclose(orig_llama_embed, llama_embed)

    llama_rope = llama_model.build_rope_cache(token_sample)
    llama_mask = llama_model.build_mask_cache(token_sample)
    orig_llama_mask = torch.full((1, 1, seq_len, seq_len), float("-inf"))
    orig_llama_mask = torch.triu(orig_llama_mask, diagonal=1)
    if kv_cache:
        orig_llama_block_out = orig_llama_model.layers[0](
            orig_llama_embed, 0, orig_llama_model.freqs_cis[:seq_len], orig_llama_mask
        )
        theirs_k_cache = orig_llama_model.layers[0].attention.cache_k
        theirs_v_cache = orig_llama_model.layers[0].attention.cache_v
        head_size = n_embd // n_head
        kv_cache_shape = (batch_size, n_head, block_size, head_size)
        ours_kv_cache = torch.zeros(kv_cache_shape), torch.zeros(kv_cache_shape)
        (llama_block_out, ours_kv_cache) = llama_model.transformer.h[0](
            llama_embed, llama_rope, llama_mask, seq_len, torch.arange(block_size), ours_kv_cache
        )
        ours_k_cache = ours_kv_cache[0].permute(0, 2, 1, 3)
        ours_v_cache = ours_kv_cache[1].permute(0, 2, 1, 3)
        torch.testing.assert_close(ours_k_cache, theirs_k_cache)
        torch.testing.assert_close(ours_v_cache, theirs_v_cache)
    else:
        orig_llama_block_out = orig_llama_model.layers[0](
            orig_llama_embed, 0, orig_llama_model.freqs_cis[:seq_len], orig_llama_mask
        )
        (llama_block_out, _) = llama_model.transformer.h[0](llama_embed, llama_rope, llama_mask, seq_len)
    assert torch.allclose(orig_llama_block_out, llama_block_out)

    expected = orig_llama_model(token_sample, 0)
    out = llama_model(token_sample)
    assert torch.allclose(out, expected)


@pytest.mark.skipif(not torch.cuda.is_available(), reason="Requires CUDA")
@torch.no_grad()
def test_bfloat16_llama_init(lit_llama, orig_llama) -> None:
    from lit_llama.utils import EmptyInitOnDevice

    block_size = 64
    vocab_size = 32000
    n_layer = 16
    n_head = 16
    n_embd = 32

    llama_config = lit_llama.LLaMAConfig(
        block_size=block_size, vocab_size=vocab_size, n_layer=n_layer, n_head=n_head, n_embd=n_embd
    )
    llama_model = lit_llama.LLaMA(llama_config)
    llama_model.apply(llama_model._init_weights)

    batch_size = 3

    token_sample = torch.randint(0, vocab_size, size=(batch_size, block_size), dtype=torch.int64)

    expected = llama_model(token_sample)

    with EmptyInitOnDevice(device="cuda", dtype=torch.bfloat16):
        llama_model2 = lit_llama.LLaMA(llama_config)
    llama_model2.load_state_dict(llama_model.state_dict(keep_vars=True))

    out = llama_model2(token_sample.cuda()).float().cpu()
    torch.testing.assert_close(out, expected, atol=5e-3, rtol=1e-3)


def copy_adapter_weights(llama_model, orig_llama_model) -> None:
    # copy the gating parameter
    for llama_block, orig_llama_block in zip(llama_model.transformer.h, orig_llama_model.layers):
        if hasattr(llama_block.attn, "gating_factor"):
            llama_block.attn.gating_factor.copy_(orig_llama_block.attention.gate)

    # In the original model, there is one embedding layer for all blocks combined
    orig_adapter_wte = orig_llama_model.adapter_query.weight.reshape(
        orig_llama_model.params.adapter_layer, orig_llama_model.params.adapter_len, orig_llama_model.params.dim
    )

    # In ours, the embedding layer is split across the individual attention layers
    index = 0
    for llama_block in llama_model.transformer.h:
        if hasattr(llama_block.attn, "adapter_wte"):
            llama_block.attn.adapter_wte.weight.copy_(orig_adapter_wte[index])
            index += 1


def enable_gate(model):
    for name, param in model.named_parameters():
        if "gating_factor" in name or "gate" in name:
            param.fill_(1)


@torch.no_grad()
def test_adapter_parity(orig_llama_adapter):
    """Test parity between our implementation of LLaMA-Adapter and the reference code."""
    import lit_llama.adapter as lit_llama

    orig_llama = orig_llama_adapter

    block_size = 32
    vocab_size = 100
    n_layer = 2
    n_head = 4
    n_embd = 16
    adapter_prompt_length: int = 10
    adapter_start_layer: int = 0

    llama_config = lit_llama.LLaMAConfig(
        block_size=block_size,
        vocab_size=vocab_size,
        n_layer=n_layer,
        n_head=n_head,
        n_embd=n_embd,
        adapter_prompt_length=adapter_prompt_length,
        adapter_start_layer=adapter_start_layer,
    )
    orig_llama_config = orig_llama.ModelArgs(
        dim=n_embd,
        n_layers=n_layer,
        n_heads=n_head,
        vocab_size=vocab_size,
        norm_eps=1e-5,
        max_seq_len=block_size,
        adapter_len=adapter_prompt_length,
        adapter_layer=(n_layer - adapter_start_layer),
    )

    batch_size = 3
    token_sample = torch.randint(
        0, orig_llama_config.vocab_size, size=(batch_size, orig_llama_config.max_seq_len), dtype=torch.int64
    )

    llama_model = lit_llama.LLaMA(llama_config)
    llama_model.apply(llama_model._init_weights)
    orig_llama_model = orig_llama.Transformer(orig_llama_config)

    copy_weights(llama_model, orig_llama_model)
    copy_adapter_weights(llama_model, orig_llama_model)

    # make the gate non-zero, otherwise the adapter is disabled and the model
    # identical to regular LLaMA
    enable_gate(llama_model)
    enable_gate(orig_llama_model)

    expected = orig_llama_model(token_sample, 0)
    out = llama_model(token_sample)
    assert torch.allclose(out, expected)


@pytest.mark.skipif(sys.platform in ("win32", "darwin"), reason="torch.compile not supported on this platform")
def test_model_compile(lit_llama):
    llama_config = lit_llama.LLaMAConfig(block_size=8, vocab_size=8, n_layer=2, n_head=2, n_embd=4)
    model = lit_llama.LLaMA(llama_config)
    model.apply(model._init_weights)

    model = torch.compile(model)

    sample = torch.randint(model.config.vocab_size, size=(2, model.config.block_size), dtype=torch.int64)
    for _ in range(3):
        _ = model(sample)