A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 18:15 UTC
[Model] Pipeline parallel support for JAIS (#7603)
Besher Alkurdi committed
Aug 17, 2024 at 21:11 UTC
e73f76eec6b5aa69b065923c175610d10ccabd50
2 files changed
+51
-18
vllm/config.py
+1
@@ -36,6 +36,7 @@ _PP_SUPPORTED_MODELS = [
36
"AquilaForCausalLM",
37
"DeepseekV2ForCausalLM",
38
"InternLMForCausalLM",
39
+ "JAISLMHeadModel",
40
"LlamaForCausalLM",
41
"LLaMAForCausalLM",
42
"MistralForCausalLM",
vllm/model_executor/models/jais.py
+50
-18
@@ -20,14 +20,14 @@
20
"""Inference-only Jais model compatible with HuggingFace weights."""
21
22
import math
23
-from typing import Iterable, List, Optional, Tuple
23
+from typing import Iterable, List, Optional, Tuple, Union
24
25
import torch
26
from torch import nn
27
28
from vllm.attention import Attention, AttentionMetadata
29
from vllm.config import CacheConfig
30
-from vllm.distributed import (get_tensor_model_parallel_rank,
30
+from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
31
get_tensor_model_parallel_world_size)
32
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
33
QKVParallelLinear,
@@ -43,6 +43,8 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
43
from vllm.sequence import IntermediateTensors, SamplerOutput
44
from vllm.transformers_utils.configs import JAISConfig
45
46
+from .utils import is_pp_missing_parameter, make_layers
47
+
48
49
class SwiGLUActivation(nn.Module):
50
@@ -216,6 +218,7 @@ class JAISModel(nn.Module):
218
config: JAISConfig,
219
cache_config: Optional[CacheConfig] = None,
220
quant_config: Optional[QuantizationConfig] = None,
221
+ prefix: str = "",
222
):
223
super().__init__()
224
self.config = config
@@ -231,10 +234,15 @@ class JAISModel(nn.Module):
234
self.embeddings_scale = config.embeddings_scale
235
else:
236
self.embeddings_scale = config.mup_embeddings_scale
234
- self.h = nn.ModuleList([
235
- JAISBlock(config, cache_config, quant_config)
236
- for _ in range(config.num_hidden_layers)
237
- ])
237
+
238
+ self.start_layer, self.end_layer, self.h = make_layers(
239
+ config.num_hidden_layers,
240
+ lambda prefix: JAISBlock(config=config,
241
+ cache_config=cache_config,
242
+ quant_config=quant_config),
243
+ prefix=f"{prefix}.h",
244
+ )
245
+
246
self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
247
248
def forward(
@@ -243,19 +251,29 @@ class JAISModel(nn.Module):
251
position_ids: torch.Tensor,
252
kv_caches: List[torch.Tensor],
253
attn_metadata: AttentionMetadata,
246
- ) -> torch.Tensor:
247
- inputs_embeds = self.wte(input_ids)
248
- if self.wpe is not None:
249
- position_embeds = self.wpe(position_ids)
250
- hidden_states = inputs_embeds + position_embeds
254
+ intermediate_tensors: Optional[IntermediateTensors] = None,
255
+ ) -> Union[IntermediateTensors, torch.Tensor]:
256
+ if get_pp_group().is_first_rank:
257
+ inputs_embeds = self.wte(input_ids)
258
+ if self.wpe is not None:
259
+ position_embeds = self.wpe(position_ids)
260
+ hidden_states = inputs_embeds + position_embeds
261
+ else:
262
+ hidden_states = inputs_embeds
263
+ hidden_states *= torch.tensor(float(self.embeddings_scale),
264
+ dtype=hidden_states.dtype)
265
else:
252
- hidden_states = inputs_embeds
253
- hidden_states *= torch.tensor(float(self.embeddings_scale),
254
- dtype=hidden_states.dtype)
266
+ assert intermediate_tensors is not None
267
+ hidden_states = intermediate_tensors["hidden_states"]
268
256
- for i in range(len(self.h)):
269
+ for i in range(self.start_layer, self.end_layer):
270
layer = self.h[i]
258
- hidden_states = layer(hidden_states, kv_caches[i], attn_metadata)
271
+ hidden_states = layer(hidden_states,
272
+ kv_caches[i - self.start_layer],
273
+ attn_metadata)
274
+
275
+ if not get_pp_group().is_last_rank:
276
+ return IntermediateTensors({"hidden_states": hidden_states})
277
278
hidden_states = self.ln_f(hidden_states)
279
return hidden_states
@@ -290,9 +308,9 @@ class JAISLMHeadModel(nn.Module):
308
kv_caches: List[torch.Tensor],
309
attn_metadata: AttentionMetadata,
310
intermediate_tensors: Optional[IntermediateTensors] = None,
293
- ) -> torch.Tensor:
311
+ ) -> Union[IntermediateTensors, torch.Tensor]:
312
hidden_states = self.transformer(input_ids, positions, kv_caches,
295
- attn_metadata)
313
+ attn_metadata, intermediate_tensors)
314
return hidden_states
315
316
def compute_logits(
@@ -304,6 +322,16 @@ class JAISLMHeadModel(nn.Module):
322
sampling_metadata)
323
return logits
324
325
+ def make_empty_intermediate_tensors(
326
+ self, batch_size: int, dtype: torch.dtype,
327
+ device: torch.device) -> IntermediateTensors:
328
+ return IntermediateTensors({
329
+ "hidden_states":
330
+ torch.zeros((batch_size, self.config.hidden_size),
331
+ dtype=dtype,
332
+ device=device),
333
+ })
334
+
335
def sample(
336
self,
337
logits: torch.Tensor,
@@ -327,6 +355,10 @@ class JAISLMHeadModel(nn.Module):
355
continue
356
if not name.startswith("transformer."):
357
name = "transformer." + name
358
+
359
+ if is_pp_missing_parameter(name, self):
360
+ continue
361
+
362
param = params_dict[name]
363
# The HF's GPT-2 implementation uses Conv1D instead of Linear.
364
# Because of this, we need to transpose the weights.