xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-18 07:05:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Remove Mixtral device="cuda" declarations (#4543)

Browse Source 
Remove the device="cuda" declarations in mixtral as promised in #4343
This commit is contained in:
Philipp Moritz 2024-05-01 16:30:52 -07:00 committed by GitHub
parent 6ef09b08f8
commit c9d852d601
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/mixtral.py
View File

@ -96,13 +96,11 @@ class MixtralMoE(nn.Module):
torch.empty(self.num_total_experts, torch.empty(self.num_total_experts,
2 * self.intermediate_size, 2 * self.intermediate_size,
self.hidden_size, self.hidden_size,
device="cuda",
dtype=self.params_dtype)) dtype=self.params_dtype))
self.w2s = nn.Parameter( self.w2s = nn.Parameter(
torch.empty(self.num_total_experts, torch.empty(self.num_total_experts,
self.hidden_size, self.hidden_size,
self.intermediate_size, self.intermediate_size,
device="cuda",
dtype=self.params_dtype)) dtype=self.params_dtype))
set_weight_attrs(self.ws, { set_weight_attrs(self.ws, {
@ -114,22 +112,20 @@ class MixtralMoE(nn.Module):
# Scaling factors for FP8 weights # Scaling factors for FP8 weights
self.ws_scale = nn.Parameter( self.ws_scale = nn.Parameter(
torch.ones( torch.ones(self.num_total_experts, dtype=torch.float32),
self.num_total_experts, device="cuda", dtype=torch.float32),
requires_grad=False) if self.use_fp8 else None requires_grad=False) if self.use_fp8 else None
self.w2s_scale = nn.Parameter( self.w2s_scale = nn.Parameter(
torch.ones( torch.ones(self.num_total_experts, dtype=torch.float32),
self.num_total_experts, device="cuda", dtype=torch.float32),
requires_grad=False) if self.use_fp8 else None requires_grad=False) if self.use_fp8 else None
# Scaling factors for FP8 activations # Scaling factors for FP8 activations
need_act_scales = (self.use_fp8 need_act_scales = (self.use_fp8
and quant_config.activation_scheme == "static") and quant_config.activation_scheme == "static")
self.as_scale = nn.Parameter( self.as_scale = nn.Parameter(
torch.zeros(1, device="cuda", dtype=torch.float32), torch.zeros(1, dtype=torch.float32),
requires_grad=False) if need_act_scales else None requires_grad=False) if need_act_scales else None
self.a2s_scale = nn.Parameter( self.a2s_scale = nn.Parameter(
torch.zeros(1, device="cuda", dtype=torch.float32), torch.zeros(1, dtype=torch.float32),
requires_grad=False) if need_act_scales else None requires_grad=False) if need_act_scales else None
if need_act_scales: if need_act_scales: