add gptneox embeddings, fix phi2 inputs, also fix the casting (#1083)

src/axolotl/utils/lora_embeddings.py

@@ -8,5 +8,7 @@ def get_linear_embedding_layers(model_type):
     returns the linear embedding layers needed for loras, dependent on the model arch
     """
     if model_type == "phi-msft":
-        return ["embd", "lm_head.linear"]
-    return ["lm_head", "embed_tokens"]
+        return ["embd.wte", "lm_head.linear"]
+    if model_type == "gpt_neox":
+        return ["embed_in", "embed_out"]
+    return ["embed_tokens", "lm_head"]

src/axolotl/utils/models.py

@@ -588,13 +588,14 @@ def load_model(
         log_gpu_memory_usage(LOG, "after model load", model.device)
 
     # make sure these are fp32 per Ramesh et al. (2021)
+    embedding_modules = get_linear_embedding_layers(cfg.model_config_type)
     for name, module in model.named_modules():
         if "norm" in name:
             module.to(torch.float32)
         if model_config.model_type == "btlm":
             # don't upcast lm_head for btlm
             continue
-        if "lm_head" in name or "embed_tokens" in name:
+        if any(m in name for m in embedding_modules):
             if hasattr(module, "weight"):
                 module.to(torch.float32)
 
@@ -619,15 +620,12 @@ def load_model(
 
     # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
     # convert them back to fp16/bf16 for flash-attn compatibility.
-    if needs_fa2_dtype or (
-        cfg.flash_attention
-        and (cfg.is_llama_derived_model or cfg.is_mistral_derived_model)
-    ):
+    if needs_fa2_dtype or cfg.flash_attention:
         LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
         for name, module in model.named_modules():
             if "norm" in name:
                 module.to(cfg.torch_dtype)
-            if "lm_head" in name or "embed_tokens" in name:
+            if any(m in name for m in embedding_modules):
                 if hasattr(module, "weight"):
                     module.to(cfg.torch_dtype)
 

tests/core/test_trainer_builder.py

@@ -30,6 +30,7 @@ def fixture_cfg():
             "adam_epsilon": 0.00001,
             "dataloader_num_workers": 1,
             "dataloader_pin_memory": True,
+            "model_config_type": "llama",
         }
     )
 

tests/test_validation.py

@@ -770,7 +770,7 @@ class ValidationCheckModelConfig(BaseValidation):
                 "adapter": "qlora",
                 "load_in_4bit": True,
                 "tokens": ["<|imstart|>"],
-                "lora_modules_to_save": ["embd", "lm_head.linear"],
+                "lora_modules_to_save": ["embd.wte", "lm_head.linear"],
             }
         )