mirror of
https://github.com/turboderp-org/exllamav2.git
synced 2026-04-20 14:29:28 +00:00
Fix regressions
This commit is contained in:
@@ -262,7 +262,7 @@ min_space_in_context = args.response_chunk
|
||||
# Stop conditions
|
||||
|
||||
sc = prompt_format.stop_conditions(tokenizer)
|
||||
sc = [x for x in sc if x]
|
||||
sc = [x for x in sc if x is not None]
|
||||
generator.set_stop_conditions(sc)
|
||||
|
||||
# ANSI color codes
|
||||
|
||||
@@ -401,6 +401,7 @@ class ExLlamaV2ArchParams:
|
||||
"mlp_up": ".mlp.c_fc",
|
||||
"mlp_down": ".mlp.c_proj",
|
||||
"lm_head": "model.embed_tokens",
|
||||
"norm_eps": "layer_norm_epsilon",
|
||||
})
|
||||
self.lm.mlp_act_func = "gelu"
|
||||
self.lm.norm = "layernorm"
|
||||
@@ -471,9 +472,9 @@ class ExLlamaV2ArchParams:
|
||||
"mlp_up": ".block_sparse_moe.experts.*.v1",
|
||||
"mlp_down": ".block_sparse_moe.experts.*.w2",
|
||||
"mlp_expert_gate": ".block_sparse_moe.gate",
|
||||
"lm_head": "model.embed_tokens",
|
||||
"fused_qkv": ".self_attn.Wqkv",
|
||||
})
|
||||
self.lm.norm = "layernorm"
|
||||
self.lm.is_moe = True
|
||||
|
||||
# Phi3
|
||||
|
||||
Reference in New Issue
Block a user