Implement fused modules (#747)

* MLP: Memory saving

* Remove RMSNorm restrictions

* Map packed weights to original

* FusedAttention module

* Simplify code

* Move fused modules

* Fix critical typo

* Split inplace

* Add FFT config

* Add validation of fused arguments

* Add fused arguments to config

* Update docs

* Fix validation logic

* Add fused modules to flash attn

* Only fuse during training

* Remove timing

* Formatting

* Formatting

* Formatting

* chore: lint

* chore: lint

* add e2e tests for fused llama

* no lora for tests

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -13,12 +13,18 @@ import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
 from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.llama.modeling_llama import LlamaAttention
 from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer as OriginalLlamaDecoderLayer,
 )
-from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+from transformers.models.llama.modeling_llama import (
+    LlamaMLP,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+from xformers.ops import SwiGLU
 
-from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name
 
 try:
     from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
@@ -38,6 +44,28 @@ except ImportError:
 LOG = logging.getLogger("axolotl")
 
 
+def replace_llama_mlp_with_swiglu(model):
+    for name, module in model.named_modules():
+        if isinstance(module, LlamaMLP):
+            mlp = FusedMLP(
+                module.config, module.gate_proj, module.up_proj, module.down_proj
+            )
+            set_module_name(model, name, mlp)
+
+
+def replace_llama_qkv_with_fused(model):
+    for name, module in model.named_modules():
+        if isinstance(module, LlamaAttention):
+            qkv = FusedAttention(
+                module.config,
+                module.q_proj,
+                module.k_proj,
+                module.v_proj,
+                module.o_proj,
+            )
+            set_module_name(model, name, qkv)
+
+
 def replace_llama_attn_with_flash_attn(
     packed: Optional[bool] = False,
     cross_entropy: Optional[bool] = False,
@@ -86,6 +114,91 @@ def replace_llama_attn_with_flash_attn(
             )
 
 
+class FusedAttention(LlamaAttention):
+    """
+    Fused QKV Attention layer for incrementally improved training efficiency
+    """
+
+    def __init__(
+        self,
+        config,
+        q: torch.nn.Linear,  # pylint: disable=invalid-name
+        k: torch.nn.Linear,  # pylint: disable=invalid-name
+        v: torch.nn.Linear,  # pylint: disable=invalid-name
+        o: torch.nn.Linear,  # pylint: disable=invalid-name
+    ):
+        super().__init__(config)
+        self.config = config
+        self.init_device = next(iter(q.state_dict().values())).device
+
+        # define equivalent fused qkv projection
+        self.out_features: List[int] = [q.out_features, k.out_features, v.out_features]
+        self.qkv_proj = torch.nn.Linear(
+            q.in_features, sum(self.out_features), device=self.init_device, bias=False
+        )
+        self.o_proj = o
+
+        # overwrite initialized weights with pretrained weights
+        self.qkv_proj.weight.data = torch.cat(
+            (q.weight.data, k.weight.data, v.weight.data), dim=0
+        )
+
+    def _post_training(self, model, name):
+        q_proj, k_proj, v_proj = torch.split(
+            self.qkv_proj.weight.data, self.out_features, dim=0
+        )
+
+        new_attn = LlamaAttention(self.config)
+        new_attn.q_proj.weight.data = q_proj
+        new_attn.k_proj.weight.data = k_proj
+        new_attn.v_proj.weight.data = v_proj
+
+        set_module_name(model, name, new_attn)
+
+
+class FusedMLP(torch.nn.Module):
+    """
+    Fused MLP layer for incrementally improved training efficiency
+    """
+
+    def __init__(
+        self,
+        config,
+        gate_proj: torch.nn.Linear,
+        up_proj: torch.nn.Linear,
+        down_proj: torch.nn.Linear,
+    ):
+        super().__init__()
+        self.config = config
+        self.swiglu = SwiGLU(
+            in_features=config.hidden_size,
+            hidden_features=config.intermediate_size,
+            bias=False,
+            _pack_weights=True,
+        )
+        # overwrite initialized weights with pretrained weights
+        self.swiglu.w12.weight.data = torch.cat(
+            (gate_proj.weight.data, up_proj.weight.data), dim=0
+        )
+        self.swiglu.w3.weight.data = down_proj.weight.data
+
+    def _post_training(self, model, name):
+        w1, w2 = torch.split(  # pylint: disable=invalid-name
+            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
+        )
+
+        # Assign the split weights back to the original layers
+        new_mlp = LlamaMLP(self.config)
+        new_mlp.gate_proj.weight.data = w1
+        new_mlp.up_proj.weight.data = w2
+        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
+
+        set_module_name(model, name, new_mlp)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
+        return self.swiglu(x)
+
+
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
 # requires the attention mask to be the same as the key_padding_mask
 def _prepare_decoder_attention_mask(
@@ -147,9 +260,14 @@ def flashattn_forward(
         value_states = torch.cat(value_states, dim=-1)
 
     else:
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
+        if isinstance(self, FusedAttention):
+            query_states, key_states, value_states = self.qkv_proj(hidden_states).split(
+                self.out_features, dim=-1
+            )
+        else:
+            query_states = self.q_proj(hidden_states)
+            key_states = self.k_proj(hidden_states)
+            value_states = self.v_proj(hidden_states)
 
     query_states = query_states.view(
         bsz, q_len, self.num_heads, self.head_dim

src/axolotl/monkeypatch/utils.py

@@ -101,3 +101,16 @@ def get_cu_seqlens_from_pos_ids(position_ids):
         max_seq_lens.append(max_seq_len)
 
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
+
+
+def set_module_name(model, name, value):
+    if "." in name:
+        parent_name = name.rsplit(".", 1)[0]
+        child_name = name[len(parent_name) + 1 :]
+        parent = model.get_submodule(parent_name)
+    else:
+        parent_name = ""
+        parent = model
+        child_name = name
+
+    setattr(parent, child_name, value)

src/axolotl/train.py

@@ -40,10 +40,7 @@ class TrainDatasetMeta:
 
 
 def train(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-    dataset_meta: TrainDatasetMeta,
+    *, cfg: DictDefault, cli_args: TrainerCliArgs, dataset_meta: TrainDatasetMeta
 ):
     # load the tokenizer first
     LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
@@ -120,6 +117,11 @@ def train(
 
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
+    # post training
+    for name, module in model.named_modules():
+        if hasattr(module, "_post_training"):
+            module._post_training(model, name)  # pylint: disable=protected-access
+
     if trainer.is_fsdp_enabled:
         trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT")
         LOG.info("Set FSDP state dict type to FULL_STATE_DICT for saving.")

src/axolotl/utils/config.py

@@ -189,9 +189,15 @@ def validate_config(cfg):
             if not cfg.load_in_4bit:
                 raise ValueError("Require cfg.load_in_4bit to be True for qlora")
 
+        if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
+            raise ValueError("Fused modules are not supported with QLoRA")
+
     if not cfg.load_in_8bit and cfg.adapter == "lora":
         LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
 
+    if cfg.adapter == "lora" and (cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp):
+        raise ValueError("Fused modules are not supported with LoRA")
+
     if cfg.relora_steps:
         if cfg.adapter not in ("lora", "qlora"):
             raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
@@ -205,6 +211,9 @@ def validate_config(cfg):
         if cfg.lr_scheduler == "one_cycle":
             raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
 
+        if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
+            raise ValueError("Fused modules are not supported with ReLoRA")
+
     if cfg.trust_remote_code:
         LOG.warning(
             "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."

src/axolotl/utils/models.py

@@ -272,6 +272,20 @@ def load_model(
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
             )
+
+            if cfg.flash_attention and not inference:
+                from axolotl.monkeypatch.llama_attn_hijack_flash import (
+                    replace_llama_mlp_with_swiglu,
+                    replace_llama_qkv_with_fused,
+                )
+
+                if cfg.flash_attn_fuse_mlp:
+                    LOG.info("patching with SwiGLU")
+                    replace_llama_mlp_with_swiglu(model)
+
+                if cfg.flash_attn_fuse_qkv:
+                    LOG.info("patching with fused QKV")
+                    replace_llama_qkv_with_fused(model)
         # elif model_type == "GPTNeoXForCausalLM" and cfg.flash_attention:
         #     This is a WIP, still an issue with the backward pass
         #     RuntimeError: grad can be implicitly created only for scalar outputs