removing deepspeed guard for LoRA Triton kernels
This commit is contained in:
@@ -1224,17 +1224,12 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
|
||||
):
|
||||
capabilities = data.get("capabilities")
|
||||
is_fsdp = data.get("fsdp") is not None
|
||||
is_deepspeed = data.get("deepspeed") is not None
|
||||
|
||||
if capabilities and capabilities.get("n_gpu", 0) > 1:
|
||||
if is_fsdp:
|
||||
raise ValueError(
|
||||
"lora_mlp_kernel, lora_qkv_kernel, and lora_o_kernel are not compatible with FSDP."
|
||||
)
|
||||
if is_deepspeed:
|
||||
raise ValueError(
|
||||
"lora_mlp_kernel, lora_qkv_kernel, and lora_o_kernel are not compatible with DeepSpeed."
|
||||
)
|
||||
return data
|
||||
|
||||
@model_validator(mode="before")
|
||||
|
||||
Reference in New Issue
Block a user