Update recast1B_llama/modeling_recast_llama.py
Browse files
recast1B_llama/modeling_recast_llama.py
CHANGED
@@ -32,35 +32,33 @@ class MLPTemplateBank(nn.Module):
|
|
32 |
self.coef_shape = (coef_rows, coef_columns)
|
33 |
|
34 |
assert coef_columns is not None, "coef_columns must not be None"
|
35 |
-
|
36 |
# Ensure divisibility for proper reshaping
|
37 |
-
assert (
|
38 |
-
|
39 |
-
|
|
|
40 |
template_size = self.hidden_size * self.intermediate_size // coef_rows
|
41 |
-
|
42 |
-
self.up_templates = nn.Parameter(
|
43 |
-
|
44 |
-
|
45 |
-
self.gate_templates = nn.Parameter(
|
46 |
-
torch.randn(coef_columns, template_size)
|
47 |
-
)
|
48 |
-
|
49 |
# Better initialization
|
50 |
nn.init.xavier_uniform_(self.up_templates)
|
51 |
nn.init.xavier_uniform_(self.gate_templates)
|
52 |
|
53 |
def forward(self, up_coeffs, gate_coeffs):
|
54 |
# Compute chunked weights
|
55 |
-
up_chunks = torch.matmul(up_coeffs, self.up_templates)
|
56 |
gate_chunks = torch.matmul(gate_coeffs, self.gate_templates)
|
57 |
-
|
58 |
# Reshape to final weight matrices
|
59 |
up_weights = up_chunks.reshape(self.intermediate_size, self.hidden_size)
|
60 |
gate_weights = gate_chunks.reshape(self.intermediate_size, self.hidden_size)
|
61 |
-
|
62 |
return up_weights, gate_weights
|
63 |
|
|
|
64 |
class SharedLlamaMLP(nn.Module):
|
65 |
def __init__(self, config, bank):
|
66 |
super().__init__()
|
@@ -68,7 +66,9 @@ class SharedLlamaMLP(nn.Module):
|
|
68 |
self.bank = bank
|
69 |
self.hidden_size = config.hidden_size
|
70 |
self.intermediate_size = config.intermediate_size
|
71 |
-
self.down_proj = nn.Linear(
|
|
|
|
|
72 |
|
73 |
# Initialize coefficients with proper shapes
|
74 |
self.up_coefficients = nn.Parameter(torch.randn(bank.coef_shape))
|
@@ -90,31 +90,37 @@ class SharedLlamaMLP(nn.Module):
|
|
90 |
def forward(self, x):
|
91 |
# Generate weights using template bank
|
92 |
up_weights, gate_weights = self.bank(
|
93 |
-
self.up_coefficients,
|
94 |
-
self.gate_coefficients # Fixed order
|
95 |
)
|
96 |
-
|
97 |
# Apply SwiGLU: SiLU(gate * x) * up * x
|
98 |
-
hidden_states = self.act_fn(
|
|
|
|
|
99 |
output = self.down_proj(hidden_states)
|
100 |
|
101 |
return output
|
102 |
|
|
|
103 |
class AttTemplateBank(nn.Module):
|
104 |
def __init__(self, config, coef_rows, coef_columns):
|
105 |
super().__init__()
|
106 |
self.hidden_size = config.hidden_size
|
107 |
self.num_heads = config.num_attention_heads
|
108 |
self.head_dim = config.hidden_size // config.num_attention_heads
|
109 |
-
self.num_key_value_heads = getattr(
|
|
|
|
|
110 |
self.kv_dim = self.num_key_value_heads * self.head_dim
|
111 |
self.coef_shape = (coef_rows, coef_columns)
|
112 |
|
113 |
# Ensure divisibility
|
114 |
-
assert (
|
115 |
-
|
116 |
-
|
117 |
-
|
|
|
|
|
118 |
|
119 |
# Create templates for Q, K, V
|
120 |
self.q_templates = nn.Parameter(
|
@@ -144,9 +150,15 @@ class AttTemplateBank(nn.Module):
|
|
144 |
v_weights = v_chunks.reshape(self.kv_dim, self.hidden_size)
|
145 |
|
146 |
return q_weights, k_weights, v_weights
|
147 |
-
|
|
|
148 |
class SharedLlamaAttention(nn.Module):
|
149 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
150 |
super().__init__()
|
151 |
self.config = config
|
152 |
self.bank = bank
|
@@ -155,15 +167,21 @@ class SharedLlamaAttention(nn.Module):
|
|
155 |
self.hidden_size = config.hidden_size
|
156 |
self.num_heads = config.num_attention_heads
|
157 |
self.head_dim = self.hidden_size // self.num_heads
|
158 |
-
self.num_key_value_heads = getattr(
|
|
|
|
|
159 |
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
160 |
self.max_position_embeddings = config.max_position_embeddings
|
161 |
-
self.rope_theta = getattr(config,
|
162 |
self.is_causal = True
|
163 |
-
|
164 |
-
self.o_proj = nn.Linear(
|
|
|
|
|
|
|
|
|
165 |
self.rotary_emb = LlamaRotaryEmbedding(config=self.config)
|
166 |
-
|
167 |
# Initialize coefficients with proper shapes
|
168 |
self.q_coefficients = nn.Parameter(torch.randn(bank.coef_shape))
|
169 |
self.k_coefficients = nn.Parameter(torch.randn(bank.coef_shape))
|
@@ -187,50 +205,64 @@ class SharedLlamaAttention(nn.Module):
|
|
187 |
**kwargs,
|
188 |
):
|
189 |
bsz, q_len, _ = hidden_states.size()
|
190 |
-
|
191 |
# Generate weights using template bank
|
192 |
-
q_weights, k_weights, v_weights = self.bank(
|
193 |
-
self.q_coefficients,
|
194 |
-
self.k_coefficients,
|
195 |
-
self.v_coefficients
|
196 |
)
|
197 |
|
198 |
# Apply projections
|
199 |
query_states = F.linear(hidden_states, q_weights)
|
200 |
key_states = F.linear(hidden_states, k_weights)
|
201 |
value_states = F.linear(hidden_states, v_weights)
|
202 |
-
|
203 |
# Reshape for multi-head attention
|
204 |
-
query_states = query_states.view(
|
205 |
-
|
206 |
-
|
207 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
208 |
# Apply rotary embeddings
|
209 |
if position_embeddings is None:
|
210 |
cos, sin = self.rotary_emb(value_states, position_ids)
|
211 |
else:
|
212 |
cos, sin = position_embeddings
|
213 |
-
query_states, key_states = apply_rotary_pos_emb(
|
214 |
-
|
|
|
|
|
215 |
# Handle past key values
|
216 |
if past_key_value is not None:
|
217 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
218 |
-
key_states, value_states = past_key_value.update(
|
219 |
-
|
|
|
|
|
220 |
# Repeat key/value for grouped query attention
|
221 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
222 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
223 |
|
224 |
# Compute attention
|
225 |
-
attn_weights = torch.matmul(
|
|
|
|
|
226 |
|
227 |
if attention_mask is not None:
|
228 |
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
229 |
attn_weights = attn_weights + causal_mask
|
230 |
|
231 |
# Apply softmax and dropout
|
232 |
-
attn_weights = nn.functional.softmax(
|
233 |
-
|
|
|
|
|
|
|
|
|
234 |
attn_output = torch.matmul(attn_weights, value_states)
|
235 |
|
236 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
@@ -242,10 +274,10 @@ class SharedLlamaAttention(nn.Module):
|
|
242 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
243 |
attn_output = attn_output.reshape(bsz, q_len, -1)
|
244 |
attn_output = self.o_proj(attn_output)
|
245 |
-
|
246 |
if not output_attentions:
|
247 |
attn_weights = None
|
248 |
-
|
249 |
return attn_output, attn_weights, past_key_value
|
250 |
|
251 |
|
@@ -269,6 +301,8 @@ class RECAST1B_llamaModel(PreTrainedModel):
|
|
269 |
config_class = RECAST1B_llama
|
270 |
base_model_prefix = "llama"
|
271 |
supports_gradient_checkpointing = True
|
|
|
|
|
272 |
|
273 |
def __init__(self, config):
|
274 |
super().__init__(config)
|
@@ -641,6 +675,8 @@ class RECAST1B_LlamaForCausalLM(PreTrainedModel, GenerationMixin):
|
|
641 |
config_class = RECAST1B_llama
|
642 |
base_model_prefix = "llama"
|
643 |
supports_gradient_checkpointing = True
|
|
|
|
|
644 |
|
645 |
def __init__(self, config):
|
646 |
super().__init__(config)
|
|
|
32 |
self.coef_shape = (coef_rows, coef_columns)
|
33 |
|
34 |
assert coef_columns is not None, "coef_columns must not be None"
|
35 |
+
|
36 |
# Ensure divisibility for proper reshaping
|
37 |
+
assert (
|
38 |
+
self.hidden_size * self.intermediate_size
|
39 |
+
) % coef_rows == 0, f"hidden_size * intermediate_size ({self.hidden_size * self.intermediate_size}) must be divisible by coef_rows ({coef_rows})"
|
40 |
+
|
41 |
template_size = self.hidden_size * self.intermediate_size // coef_rows
|
42 |
+
|
43 |
+
self.up_templates = nn.Parameter(torch.randn(coef_columns, template_size))
|
44 |
+
self.gate_templates = nn.Parameter(torch.randn(coef_columns, template_size))
|
45 |
+
|
|
|
|
|
|
|
|
|
46 |
# Better initialization
|
47 |
nn.init.xavier_uniform_(self.up_templates)
|
48 |
nn.init.xavier_uniform_(self.gate_templates)
|
49 |
|
50 |
def forward(self, up_coeffs, gate_coeffs):
|
51 |
# Compute chunked weights
|
52 |
+
up_chunks = torch.matmul(up_coeffs, self.up_templates)
|
53 |
gate_chunks = torch.matmul(gate_coeffs, self.gate_templates)
|
54 |
+
|
55 |
# Reshape to final weight matrices
|
56 |
up_weights = up_chunks.reshape(self.intermediate_size, self.hidden_size)
|
57 |
gate_weights = gate_chunks.reshape(self.intermediate_size, self.hidden_size)
|
58 |
+
|
59 |
return up_weights, gate_weights
|
60 |
|
61 |
+
|
62 |
class SharedLlamaMLP(nn.Module):
|
63 |
def __init__(self, config, bank):
|
64 |
super().__init__()
|
|
|
66 |
self.bank = bank
|
67 |
self.hidden_size = config.hidden_size
|
68 |
self.intermediate_size = config.intermediate_size
|
69 |
+
self.down_proj = nn.Linear(
|
70 |
+
config.intermediate_size, config.hidden_size, bias=False
|
71 |
+
)
|
72 |
|
73 |
# Initialize coefficients with proper shapes
|
74 |
self.up_coefficients = nn.Parameter(torch.randn(bank.coef_shape))
|
|
|
90 |
def forward(self, x):
|
91 |
# Generate weights using template bank
|
92 |
up_weights, gate_weights = self.bank(
|
93 |
+
self.up_coefficients, self.gate_coefficients # Fixed order
|
|
|
94 |
)
|
95 |
+
|
96 |
# Apply SwiGLU: SiLU(gate * x) * up * x
|
97 |
+
hidden_states = self.act_fn(
|
98 |
+
F.linear(x, gate_weights, self.gate_bias)
|
99 |
+
) * F.linear(x, up_weights, self.up_bias)
|
100 |
output = self.down_proj(hidden_states)
|
101 |
|
102 |
return output
|
103 |
|
104 |
+
|
105 |
class AttTemplateBank(nn.Module):
|
106 |
def __init__(self, config, coef_rows, coef_columns):
|
107 |
super().__init__()
|
108 |
self.hidden_size = config.hidden_size
|
109 |
self.num_heads = config.num_attention_heads
|
110 |
self.head_dim = config.hidden_size // config.num_attention_heads
|
111 |
+
self.num_key_value_heads = getattr(
|
112 |
+
config, "num_key_value_heads", config.num_attention_heads
|
113 |
+
)
|
114 |
self.kv_dim = self.num_key_value_heads * self.head_dim
|
115 |
self.coef_shape = (coef_rows, coef_columns)
|
116 |
|
117 |
# Ensure divisibility
|
118 |
+
assert (
|
119 |
+
self.hidden_size * self.hidden_size
|
120 |
+
) % coef_rows == 0, "Q projection size must be divisible by coef_rows"
|
121 |
+
assert (
|
122 |
+
self.kv_dim * self.hidden_size
|
123 |
+
) % coef_rows == 0, "K/V projection size must be divisible by coef_rows"
|
124 |
|
125 |
# Create templates for Q, K, V
|
126 |
self.q_templates = nn.Parameter(
|
|
|
150 |
v_weights = v_chunks.reshape(self.kv_dim, self.hidden_size)
|
151 |
|
152 |
return q_weights, k_weights, v_weights
|
153 |
+
|
154 |
+
|
155 |
class SharedLlamaAttention(nn.Module):
|
156 |
+
def __init__(
|
157 |
+
self,
|
158 |
+
config,
|
159 |
+
layer_idx: Optional[int] = None,
|
160 |
+
bank: Optional[AttTemplateBank] = None,
|
161 |
+
):
|
162 |
super().__init__()
|
163 |
self.config = config
|
164 |
self.bank = bank
|
|
|
167 |
self.hidden_size = config.hidden_size
|
168 |
self.num_heads = config.num_attention_heads
|
169 |
self.head_dim = self.hidden_size // self.num_heads
|
170 |
+
self.num_key_value_heads = getattr(
|
171 |
+
config, "num_key_value_heads", config.num_attention_heads
|
172 |
+
)
|
173 |
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
174 |
self.max_position_embeddings = config.max_position_embeddings
|
175 |
+
self.rope_theta = getattr(config, "rope_theta", 10000.0)
|
176 |
self.is_causal = True
|
177 |
+
|
178 |
+
self.o_proj = nn.Linear(
|
179 |
+
self.hidden_size,
|
180 |
+
self.hidden_size,
|
181 |
+
bias=getattr(config, "attention_bias", False),
|
182 |
+
)
|
183 |
self.rotary_emb = LlamaRotaryEmbedding(config=self.config)
|
184 |
+
|
185 |
# Initialize coefficients with proper shapes
|
186 |
self.q_coefficients = nn.Parameter(torch.randn(bank.coef_shape))
|
187 |
self.k_coefficients = nn.Parameter(torch.randn(bank.coef_shape))
|
|
|
205 |
**kwargs,
|
206 |
):
|
207 |
bsz, q_len, _ = hidden_states.size()
|
208 |
+
|
209 |
# Generate weights using template bank
|
210 |
+
q_weights, k_weights, v_weights = self.bank(
|
211 |
+
self.q_coefficients, self.k_coefficients, self.v_coefficients
|
|
|
|
|
212 |
)
|
213 |
|
214 |
# Apply projections
|
215 |
query_states = F.linear(hidden_states, q_weights)
|
216 |
key_states = F.linear(hidden_states, k_weights)
|
217 |
value_states = F.linear(hidden_states, v_weights)
|
218 |
+
|
219 |
# Reshape for multi-head attention
|
220 |
+
query_states = query_states.view(
|
221 |
+
bsz, q_len, self.num_heads, self.head_dim
|
222 |
+
).transpose(1, 2)
|
223 |
+
key_states = key_states.view(
|
224 |
+
bsz, q_len, self.num_key_value_heads, self.head_dim
|
225 |
+
).transpose(1, 2)
|
226 |
+
value_states = value_states.view(
|
227 |
+
bsz, q_len, self.num_key_value_heads, self.head_dim
|
228 |
+
).transpose(1, 2)
|
229 |
+
|
230 |
# Apply rotary embeddings
|
231 |
if position_embeddings is None:
|
232 |
cos, sin = self.rotary_emb(value_states, position_ids)
|
233 |
else:
|
234 |
cos, sin = position_embeddings
|
235 |
+
query_states, key_states = apply_rotary_pos_emb(
|
236 |
+
query_states, key_states, cos, sin
|
237 |
+
)
|
238 |
+
|
239 |
# Handle past key values
|
240 |
if past_key_value is not None:
|
241 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
242 |
+
key_states, value_states = past_key_value.update(
|
243 |
+
key_states, value_states, self.layer_idx, cache_kwargs
|
244 |
+
)
|
245 |
+
|
246 |
# Repeat key/value for grouped query attention
|
247 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
248 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
249 |
|
250 |
# Compute attention
|
251 |
+
attn_weights = torch.matmul(
|
252 |
+
query_states, key_states.transpose(2, 3)
|
253 |
+
) / math.sqrt(self.head_dim)
|
254 |
|
255 |
if attention_mask is not None:
|
256 |
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
257 |
attn_weights = attn_weights + causal_mask
|
258 |
|
259 |
# Apply softmax and dropout
|
260 |
+
attn_weights = nn.functional.softmax(
|
261 |
+
attn_weights, dim=-1, dtype=torch.float32
|
262 |
+
).to(query_states.dtype)
|
263 |
+
attn_weights = nn.functional.dropout(
|
264 |
+
attn_weights, p=self.attention_dropout, training=self.training
|
265 |
+
)
|
266 |
attn_output = torch.matmul(attn_weights, value_states)
|
267 |
|
268 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
|
|
274 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
275 |
attn_output = attn_output.reshape(bsz, q_len, -1)
|
276 |
attn_output = self.o_proj(attn_output)
|
277 |
+
|
278 |
if not output_attentions:
|
279 |
attn_weights = None
|
280 |
+
|
281 |
return attn_output, attn_weights, past_key_value
|
282 |
|
283 |
|
|
|
301 |
config_class = RECAST1B_llama
|
302 |
base_model_prefix = "llama"
|
303 |
supports_gradient_checkpointing = True
|
304 |
+
_no_split_modules = ["LlamaDecoderLayer"] # Add this
|
305 |
+
_skip_keys_device_placement = "past_key_values" # Add this
|
306 |
|
307 |
def __init__(self, config):
|
308 |
super().__init__(config)
|
|
|
675 |
config_class = RECAST1B_llama
|
676 |
base_model_prefix = "llama"
|
677 |
supports_gradient_checkpointing = True
|
678 |
+
_no_split_modules = ["LlamaDecoderLayer"] # Add this
|
679 |
+
_skip_keys_device_placement = "past_key_values" # Add this
|
680 |
|
681 |
def __init__(self, config):
|
682 |
super().__init__(config)
|