[Feature][Kernel] Support bitsandbytes quantization and QLoRA (#4776)
This commit is contained in:
@@ -319,6 +319,14 @@ class LlamaForCausalLM(nn.Module):
|
||||
"lm_head": "output_embeddings",
|
||||
}
|
||||
embedding_padding_modules = ["lm_head"]
|
||||
bitsandbytes_stacked_params_mapping = {
|
||||
# shard_name, weight_name, index
|
||||
"q_proj": ("qkv_proj", 0),
|
||||
"k_proj": ("qkv_proj", 1),
|
||||
"v_proj": ("qkv_proj", 2),
|
||||
"gate_proj": ("gate_up_proj", 0),
|
||||
"up_proj": ("gate_up_proj", 1),
|
||||
}
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user