@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / e7cfc4ef4c
Fork 0

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, 05:45 UTC

[Interleaved ATTN] Support for Mistral-8B (#10591)

Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: youkaichao <youkaichao@gmail.com>

Patrick von Platen committed Nov 30, 2024 at 08:45 UTC e7cfc4ef4cc017e0a0229adff9f4b143b38fb421
1 file changed +15 -1
vllm/model_executor/models/llama.py
+15 -1
@@ -54,7 +54,7 @@ from vllm.sequence import IntermediateTensors, PoolerOutput
54
55 from .interfaces import SupportsLoRA, SupportsPP
56 from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
57 - is_pp_missing_parameter,
57 + extract_layer_index, is_pp_missing_parameter,
58 make_empty_intermediate_tensors_factory, make_layers,
59 maybe_prefix)
60
@@ -114,6 +114,7 @@ class LlamaAttention(nn.Module):
114 prefix: str = "",
115 ) -> None:
116 super().__init__()
117 + layer_idx = extract_layer_index(prefix)
118 self.hidden_size = hidden_size
119 tp_size = get_tensor_model_parallel_world_size()
120 self.total_num_heads = num_heads
@@ -168,6 +169,18 @@ class LlamaAttention(nn.Module):
169 rope_scaling=rope_scaling,
170 is_neox_style=is_neox_style,
171 )
172 +
173 + if hasattr(config, "interleaved_sliding_window"):
174 + if isinstance(config.interleaved_sliding_window, int):
175 + sliding_window = config.interleaved_sliding_window
176 + elif isinstance(config.interleaved_sliding_window, list):
177 + sw_idx = layer_idx % len(config.interleaved_sliding_window)
178 + sliding_window = config.interleaved_sliding_window[sw_idx]
179 + else:
180 + raise ValueError(f"{type(sliding_window)} is not supported.")
181 + else:
182 + sliding_window = None
183 +
184 self.attn = Attention(
185 self.num_heads,
186 self.head_dim,
@@ -175,6 +188,7 @@ class LlamaAttention(nn.Module):
188 num_kv_heads=self.num_kv_heads,
189 cache_config=cache_config,
190 quant_config=quant_config,
191 + per_layer_sliding_window=sliding_window,
192 prefix=f"{prefix}.attn",
193 )
194