zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-09-01 12:49:50 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Kv cache (#643)

Browse Source 
* in place kv_cache

* fix

* fix kv cache size

* partially fix kv cache dtype

* step kv cache

* multiple of step size

* more teests + kv cache

* more kv cache

* udpate all models to use kv cache
This commit is contained in:
Awni Hannun
2024-05-08 08:18:13 -07:00
committed by GitHub
parent bfbc0e434a
commit ee60e2a9d5
22 changed files with 534 additions and 298 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
llms/mlx_lm/models/llama.py
View File

@@ -78,11 +78,9 @@ class Attention(nn.Module):
values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
if cache is not None:
key_cache, value_cache = cache
queries = self.rope(queries, offset=key_cache.shape[2])
keys = self.rope(keys, offset=key_cache.shape[2])
keys = mx.concatenate([key_cache, keys], axis=2)
values = mx.concatenate([value_cache, values], axis=2)
queries = self.rope(queries, offset=cache.offset)
keys = self.rope(keys, offset=cache.offset)
keys, values = cache.update_and_fetch(keys, values)
else:
queries = self.rope(queries)
keys = self.rope(keys)
@@ -91,7 +89,7 @@ class Attention(nn.Module):
queries, keys, values, scale=self.scale, mask=mask
)
output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
return self.o_proj(output), (keys, values)
return self.o_proj(output)
class MLP(nn.Module):
@@ -124,11 +122,11 @@ class TransformerBlock(nn.Module):
mask: Optional[mx.array] = None,
cache: Optional[Tuple[mx.array, mx.array]] = None,
) -> mx.array:
r, cache = self.self_attn(self.input_layernorm(x), mask, cache)
r = self.self_attn(self.input_layernorm(x), mask, cache)
h = x + r
r = self.mlp(self.post_attention_layernorm(h))
out = h + r
return out, cache
return out
def create_additive_causal_mask(N: int, offset: int = 0):
@@ -168,10 +166,10 @@ class LlamaModel(nn.Module):
if cache is None:
cache = [None] * len(self.layers)
for e, layer in enumerate(self.layers):
h, cache[e] = layer(h, mask, cache[e])
for layer, c in zip(self.layers, cache):
h = layer(h, mask, cache=c)
return self.norm(h), cache
return self.norm(h)
class Model(nn.Module):
@@ -180,14 +178,15 @@ class Model(nn.Module):
self.model_type = args.model_type
self.model = LlamaModel(args)
self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
self.args = args
def __call__(
self,
inputs: mx.array,
cache=None,
):
out, cache = self.model(inputs, cache)
return self.lm_head(out), cache
out = self.model(inputs, cache)
return self.lm_head(out)
def sanitize(self, weights):
# Remove unused precomputed rotary freqs
@@ -198,3 +197,11 @@ class Model(nn.Module):
@property
def layers(self):
return self.model.layers
@property
def head_dim(self):
return self.args.hidden_size // self.args.num_attention_heads
@property
def n_kv_heads(self):
return self.args.num_key_value_heads