ViperEkura
0c1b7664c1
- Eliminate core/ directory into cache/, runtime/, network/ subpackages plus flat modules
- Split cache.py (647 lines) into cache/{buffer,strategy,pool}.py by layer
- Add explicit ContiguousStrategy, make AllocationStrategy a real ABC
- Move TaskCacheState to cache/strategy.py, drop string forward references
- Rename api/ to network/, server.py to app.py
- Move sample.py into runtime/ alongside executor and graph
- Simplify TaskCacheManager.__init__ to single pool param
- Expose pool.strategy and pool.req_pool as public properties
- Fix KVCache import in attention_backend.py (TYPE_CHECKING guard)
- Fix steady-state decode reading uninitialized position_ids on first step
140 lines
4.6 KiB
Python
140 lines
4.6 KiB
Python
from typing import Any, Dict, Mapping, Optional
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
from torch import Tensor
|
|
|
|
from astrai.config.model_config import AutoRegressiveLMConfig
|
|
from astrai.inference.cache import KVCache
|
|
from astrai.model.automodel import AutoModel, ModelFactory
|
|
from astrai.model.components.decoder_block import DecoderBlock
|
|
from astrai.model.components.embedding import Embedding
|
|
from astrai.model.components.linear import Linear
|
|
from astrai.model.components.norm import RMSNorm
|
|
from astrai.model.components.rope import RotaryEmbedding
|
|
|
|
|
|
def process_attention_mask(
|
|
input_mask: Optional[Tensor],
|
|
) -> Optional[Tensor]:
|
|
if input_mask is None:
|
|
return None
|
|
if input_mask.dim() == 2:
|
|
return input_mask[:, None, None, :]
|
|
if input_mask.dim() == 3:
|
|
return input_mask[:, None, :, :]
|
|
return input_mask
|
|
|
|
|
|
@ModelFactory.register("autoregressive_lm")
|
|
class AutoRegressiveLM(AutoModel):
|
|
"""Autoregressive language model with paged KV cache."""
|
|
|
|
def __init__(self, config: AutoRegressiveLMConfig):
|
|
super().__init__(config)
|
|
self.config = config
|
|
rope_dim = (
|
|
config.qk_rope_head_dim
|
|
if config.attn_type == "mla"
|
|
else config.hidden_size // config.num_attention_heads
|
|
)
|
|
rope_base = config.rope_theta if config.rope_theta is not None else 10000
|
|
self.rotary_embedding = RotaryEmbedding(
|
|
rope_dim,
|
|
config.max_position_embeddings,
|
|
rope_base,
|
|
rope_scaling=config.rope_scaling,
|
|
)
|
|
self.embed_tokens = Embedding(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
neftune_alpha=config.neftune_alpha,
|
|
)
|
|
|
|
self.layers = nn.ModuleList(
|
|
[
|
|
DecoderBlock(config, layer_id)
|
|
for layer_id in range(config.num_hidden_layers)
|
|
]
|
|
)
|
|
|
|
self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
|
|
self.lm_head = Linear(config.hidden_size, config.vocab_size)
|
|
|
|
if self.config.tie_word_embeddings is True:
|
|
self.lm_head.weight = self.embed_tokens.weight
|
|
|
|
self.apply(self._init_weights)
|
|
|
|
def _init_weights(self, module):
|
|
if hasattr(module, "reset_parameters"):
|
|
module.reset_parameters()
|
|
|
|
def load_state_dict(self, state_dict: Mapping[str, Any], strict=True, assign=False):
|
|
lm_head_key = "lm_head.weight"
|
|
embed_key = "embed_tokens.weight"
|
|
|
|
state_dict = dict(state_dict)
|
|
|
|
if self.config.tie_word_embeddings is True:
|
|
# same tensor for embed and lm_head
|
|
if embed_key in state_dict:
|
|
state_dict[lm_head_key] = state_dict[embed_key]
|
|
else:
|
|
if lm_head_key not in state_dict and embed_key in state_dict:
|
|
# clone to avoid sharing gradients
|
|
state_dict[lm_head_key] = torch.clone(state_dict[embed_key])
|
|
|
|
return super().load_state_dict(state_dict, strict, assign)
|
|
|
|
def state_dict(self, destination=None, prefix="", keep_vars=False):
|
|
state_dict = super().state_dict(
|
|
destination=destination, prefix=prefix, keep_vars=keep_vars
|
|
)
|
|
|
|
if self.config.tie_word_embeddings is True:
|
|
lm_head_key = prefix + "lm_head.weight"
|
|
if lm_head_key in state_dict:
|
|
del state_dict[lm_head_key]
|
|
|
|
return state_dict
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: Tensor,
|
|
input_mask: Optional[Tensor] = None,
|
|
kv_cache: Optional[KVCache] = None,
|
|
position_ids: Optional[Tensor] = None,
|
|
) -> Dict[str, Tensor]:
|
|
assert input_ids.ndim == 2
|
|
|
|
x = self.embed_tokens(input_ids)
|
|
rotary_emb = self.rotary_embedding(x, position_ids)
|
|
attn_mask = process_attention_mask(input_mask)
|
|
use_sdpa_causal_mask = attn_mask is None
|
|
|
|
aux_losses = []
|
|
router_stats_list = []
|
|
for layer in self.layers:
|
|
layer_output = layer(
|
|
x,
|
|
rotary_emb,
|
|
attn_mask,
|
|
kv_cache,
|
|
use_sdpa_causal_mask,
|
|
)
|
|
x = layer_output["hidden_states"]
|
|
stats = layer_output.get("router_stats")
|
|
if stats is not None:
|
|
aux_losses.append(layer_output["aux_loss"])
|
|
router_stats_list.append(stats)
|
|
|
|
hidden_states = self.norm(x)
|
|
logits = self.lm_head(hidden_states)
|
|
|
|
output = {"logits": logits, "hidden_states": hidden_states}
|
|
if aux_losses:
|
|
output["aux_loss"] = torch.stack(aux_losses).mean()
|
|
output["router_stats"] = router_stats_list
|
|
return output
|