@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / cdc72e3c80
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, 10:45 UTC

[Model] Remap FP8 kv_scale in CommandR and DBRX (#9174)

Hui Liu committed Oct 8, 2024 at 23:43 UTC cdc72e3c80b7029c49de9667150f68481f386956
2 files changed +14 -2
vllm/model_executor/models/commandr.py
+7 -1
@@ -41,7 +41,8 @@ from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
41 from vllm.model_executor.layers.vocab_parallel_embedding import (
42 VocabParallelEmbedding)
43 from vllm.model_executor.model_loader.weight_utils import (
44 - default_weight_loader, row_parallel_weight_loader)
44 + default_weight_loader, maybe_remap_kv_scale_name,
45 + row_parallel_weight_loader)
46 from vllm.model_executor.sampling_metadata import SamplingMetadata
47 from vllm.model_executor.utils import set_weight_attrs
48 from vllm.sequence import IntermediateTensors
@@ -426,6 +427,11 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
427 # Skip loading extra bias for GPTQ models.
428 if name.endswith(".bias") and name not in params_dict:
429 continue
430 + # Remapping the name of FP8 kv-scale.
431 + name = maybe_remap_kv_scale_name(name, params_dict)
432 + if name is None:
433 + continue
434 +
435 if is_pp_missing_parameter(name, self):
436 continue
437 param = params_dict[name]
vllm/model_executor/models/dbrx.py
+7 -1
@@ -18,7 +18,8 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
18 from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
19 from vllm.model_executor.layers.vocab_parallel_embedding import (
20 DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
21 -from vllm.model_executor.model_loader.weight_utils import default_weight_loader
21 +from vllm.model_executor.model_loader.weight_utils import (
22 + default_weight_loader, maybe_remap_kv_scale_name)
23 from vllm.model_executor.sampling_metadata import SamplingMetadata
24 from vllm.sequence import IntermediateTensors
25 from vllm.transformers_utils.configs.dbrx import DbrxConfig
@@ -425,6 +426,11 @@ class DbrxForCausalLM(nn.Module, SupportsPP):
426 weight_loader(param, loaded_weight, weight_name)
427 break
428 else:
429 + # Remapping the name of FP8 kv-scale.
430 + name = maybe_remap_kv_scale_name(name, params_dict)
431 + if name is None:
432 + continue
433 +
434 if is_pp_missing_parameter(name, self):
435 continue
436 param = params_dict[name]