Feat: Add support for upstream FA2 (axolotl-ai-cloud#626)

* Feat: Add support for upstream FA2

* chore: add is_falcon_derived_model: true to examples

* chore: add config to readme for documentation

* feat: add extra model types

* fix: remove old falcon flash patch

* chore: pin transformers and accelerate

README.md

@@ -408,6 +408,10 @@ tokenizer_legacy:
 # this is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
 
+# used to identify if the model is falcon/llama based
+is_falcon_derived_model:
+is_llama_derived_model:
+
 # whether you are training a 4-bit GPTQ quantized model
 gptq: true
 gptq_groupsize: 128 # group size

examples/falcon/config-7b-lora.yml

@@ -3,6 +3,7 @@ base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+is_falcon_derived_model: true
 load_in_8bit: true
 load_in_4bit: false
 gptq: false

examples/falcon/config-7b-qlora.yml

@@ -6,6 +6,7 @@ base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+is_falcon_derived_model: true
 load_in_8bit: false
 # enable 4bit for QLoRA
 load_in_4bit: true

examples/falcon/config-7b.yml

@@ -3,6 +3,7 @@ base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+is_falcon_derived_model: true
 load_in_8bit: false
 load_in_4bit: false
 gptq: false

requirements.txt

@@ -4,9 +4,9 @@ torch==2.0.1
 auto-gptq
 packaging
 peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git
+transformers @ git+https://github.com/huggingface/transformers.git@0ac3875011d32dc85e0e83970507e3afe8f0febb
 bitsandbytes>=0.41.1
-accelerate @ git+https://github.com/huggingface/accelerate
+accelerate @ git+https://github.com/huggingface/accelerate@80da9cfb09bb3cc9f1b385cb55d6b90d025a5fd9
 deepspeed
 addict
 evaluate

src/axolotl/utils/config.py

@@ -86,6 +86,22 @@ def normalize_config(cfg):
         or (cfg.model_type and "llama" in cfg.model_type.lower())
     )
 
+    # figure out if the model is falcon
+    cfg.is_falcon_derived_model = (
+        (
+            hasattr(model_config, "model_type")
+            and model_config.model_type
+            in [
+                "falcon",
+                "RefinedWebModel",
+                "RefinedWeb",
+            ]
+        )
+        or cfg.is_falcon_derived_model
+        or "falcon" in cfg.base_model
+        or (cfg.model_type and "rwforcausallm" in cfg.model_type.lower())
+    )
+
     log_gpu_memory_usage(LOG, "baseline", cfg.device)
 
 

src/axolotl/utils/models.py

@@ -114,25 +114,13 @@ def load_model(
 
             replace_btlm_attn_with_flash_attn(cfg.base_model)
 
-    if hasattr(model_config, "model_type") and model_config.model_type in [
-        "falcon",
-        "RefinedWebModel",
-        "RefinedWeb",
-    ]:
-        if cfg.flash_attention:
-            from axolotl.monkeypatch.falcon_attn_hijack_flash import (
-                replace_falcon_attn_with_flash_attn,
-            )
-
-            replace_falcon_attn_with_flash_attn()
-
-    if cfg.is_llama_derived_model and cfg.flash_attention:
+    if cfg.is_llama_derived_model and cfg.flash_attention and cfg.sample_packing:
         if cfg.device not in ["mps", "cpu"] and not inference:
             from axolotl.monkeypatch.llama_attn_hijack_flash import (
                 replace_llama_attn_with_flash_attn,
             )
 
-            LOG.info("patching with flash attention")
+            LOG.info("patching with flash attention for sample packing")
             replace_llama_attn_with_flash_attn(packed=cfg.sample_packing)
     elif cfg.is_llama_derived_model and cfg.xformers_attention:
         from axolotl.monkeypatch.llama_attn_hijack_xformers import (
@@ -213,6 +201,10 @@ def load_model(
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
         )
+    # sample packing uses custom FA2 patch
+    if cfg.flash_attention and not cfg.sample_packing:
+        if cfg.is_llama_derived_model or cfg.is_falcon_derived_model:
+            model_kwargs["use_flash_attention_2"] = True
     try:
         if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM