frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Change kv scaling factor by param json on nvidia gpu (#11688)

Browse Source 
Signed-off-by: bjmsong <bjmsong@126.com>
Co-authored-by: bjmsong <bjmsong@126.com>
This commit is contained in:
bjmsong
2025-01-03 05:11:39 +08:00
committed by GitHub
parent b55ed6ef8a
commit 187e32997c
5 changed files with 14 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/models/exaone.py
View File

@ -606,8 +606,9 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
# which is consistent with the practice of setting
# scaling_factor = tensor_amax / FPtype_max
scaling_factor *= 2
if hasattr(layer_self_attn, "kv_scale"):
layer_self_attn.attn._kv_scale = scaling_factor
if hasattr(layer_self_attn.attn, "_k_scale"):
layer_self_attn.attn._k_scale = scaling_factor
layer_self_attn.attn._v_scale = scaling_factor
else:
raise RuntimeError("Self attention has no KV cache scaling "
"factor attribute!")

5
vllm/model_executor/models/granite.py
View File

@ -545,8 +545,9 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
# which is consistent with the practice of setting
# scaling_factor = tensor_amax / FPtype_max
scaling_factor *= 2
if hasattr(layer_self_attn, "kv_scale"):
layer_self_attn.attn._kv_scale = scaling_factor
if hasattr(layer_self_attn.attn, "_k_scale"):
layer_self_attn.attn._k_scale = scaling_factor
layer_self_attn.attn._v_scale = scaling_factor
else:
raise RuntimeError("Self attention has no KV cache scaling "
"factor attribute!")

5
vllm/model_executor/models/llama.py
View File

@ -452,8 +452,9 @@ class LlamaModel(nn.Module):
# which is consistent with the practice of setting
# scaling_factor = tensor_amax / FPtype_max
scaling_factor *= 2
if hasattr(layer_self_attn, "kv_scale"):
layer_self_attn.attn._kv_scale = scaling_factor
if hasattr(layer_self_attn.attn, "_k_scale"):
layer_self_attn.attn._k_scale = scaling_factor
layer_self_attn.attn._v_scale = scaling_factor
else:
raise RuntimeError("Self attention has no KV cache scaling "
"factor attribute!")

5
vllm/model_executor/models/solar.py
View File

@ -565,8 +565,9 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
# which is consistent with the practice of setting
# scaling_factor = tensor_amax / FPtype_max
scaling_factor *= 2
if hasattr(layer_self_attn, "kv_scale"):
layer_self_attn.attn._kv_scale = scaling_factor
if hasattr(layer_self_attn.attn, "_k_scale"):
layer_self_attn.attn._k_scale = scaling_factor
layer_self_attn.attn._v_scale = scaling_factor
else:
raise RuntimeError("Self attention has no KV cache scaling "
"factor attribute!")

3
vllm/worker/model_runner.py
View File

@ -1136,7 +1136,8 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
self.prompt_adapter_manager.create_prompt_adapter_manager(
self.model))
if self.kv_cache_dtype == "fp8" and current_platform.is_rocm():
if self.kv_cache_dtype == "fp8" and (current_platform.is_rocm()
or current_platform.is_cuda()):
# Currently only ROCm accepts kv-cache scaling factors
# via quantization_param_path and this will be deprecated
# in the future.