Compare commits

...

4 Commits

Author SHA1 Message Date
Wing Lian
6978f09760 pre-patch the mlp 2025-07-13 23:01:49 -04:00
Wing Lian
d41b3814d0 use new patch 2025-07-13 22:40:37 -04:00
Wing Lian
1649f91cd4 wip patch 2025-07-13 22:37:18 -04:00
Wing Lian
5a063f5c75 wip state dict compatible fused mlp 2025-07-13 22:37:18 -04:00
3 changed files with 168 additions and 6 deletions

View File

@@ -398,6 +398,18 @@ class PatchManager:
"Shifted-sparse attention not currently implemented without flash attention."
)
from axolotl.monkeypatch.llama_attn_hijack_flash import (
is_xformers_swiglu_available,
)
if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
from axolotl.monkeypatch.llama_attn_hijack_flash import (
patch_mlp_with_swiglu,
)
LOG.info("Patching with SwiGLU...")
patch_mlp_with_swiglu(self.cfg.model_config_type)
def _apply_llama_flash_attn_patches(self, model):
"""Apply LLaMA-specific flash attention patches."""
if (
@@ -408,15 +420,14 @@ class PatchManager:
and not self.inference
):
# TODO(MengqingCao): split these patches seperately
from axolotl.monkeypatch.llama_attn_hijack_flash import (
is_xformers_swiglu_available,
replace_llama_mlp_with_swiglu,
from axolotl.monkeypatch.llama_attn_hijack_flash import ( # is_xformers_swiglu_available,; replace_llama_mlp_with_swiglu,
replace_llama_qkv_with_fused,
)
if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
LOG.info("Patching with SwiGLU...")
replace_llama_mlp_with_swiglu(model)
# if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
# LOG.info("Patching with SwiGLU...")
# # replace_llama_mlp_with_swiglu(model)
# patch_mlp_with_swiglu(model)
if self.cfg.flash_attn_fuse_qkv:
LOG.info("Patching with fused QKV...")

View File

@@ -82,6 +82,28 @@ def replace_llama_mlp_with_swiglu(model):
set_module_name(model, name, mlp)
def patch_mlp_with_swiglu(model_type):
if is_xformers_swiglu_available():
from axolotl.monkeypatch.xformers_ import FusedMLPv2 as FusedMLP
else:
raise RuntimeError("xformers SwiGLU not available for this environment")
try:
# Dynamically import the module and MLP class
module_path = f"transformers.models.{model_type}.modeling_{model_type}"
model_cls_prefix = "".join(
[part.capitalize() for part in model_type.split("_")]
)
module = __import__(module_path, fromlist=[f"{model_cls_prefix}MLP"])
_ = getattr(module, f"{model_cls_prefix}MLP")
setattr(module, f"{model_cls_prefix}MLP", FusedMLP)
except (ImportError, AttributeError) as e:
raise RuntimeError(
f"Could not import MLP class for model_type: {model_type}. "
f"Error: {str(e)}"
) from e
def replace_llama_qkv_with_fused(model):
for name, module in model.named_modules():
if isinstance(module, LlamaAttention):

View File

@@ -1,8 +1,11 @@
"""
Fused MLP layer for incrementally improved training efficiency
"""
from collections import OrderedDict
import torch
from torch import nn
from transformers.activations import ACT2FN
from transformers.models.llama.modeling_llama import LlamaMLP
from xformers.ops import SwiGLU
@@ -50,3 +53,129 @@ class FusedMLP(torch.nn.Module):
def forward(self, x: torch.Tensor) -> torch.Tensor: # pylint: disable=invalid-name
return self.swiglu(x)
class FusedMLPv2(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.hidden_size = config.hidden_size
self.intermediate_size = config.intermediate_size
self.swiglu = SwiGLU(
in_features=self.hidden_size,
hidden_features=self.intermediate_size,
bias=config.mlp_bias,
_pack_weights=True,
)
assert config.hidden_act == "silu"
def _convert_unpacked_to_packed_state_dict(self, unpacked_state_dict):
"""
Convert state dict from unpacked format (w1, w2, w3) to packed format (w13, w2).
"""
packed_state_dict = OrderedDict()
# Handle w1 and w3 -> w13 conversion for weights
if 'gate_proj.weight' in unpacked_state_dict and 'up_proj.weight' in unpacked_state_dict:
gate_proj_weight = unpacked_state_dict['gate_proj.weight']
up_proj_weight = unpacked_state_dict['up_proj.weight']
# Concatenate gate and up weights along output dimension (dim=0)
packed_state_dict['swiglu.w12.weight'] = torch.cat([gate_proj_weight, up_proj_weight], dim=0)
# Handle w1 and w3 -> w13 conversion for biases (if they exist)
if 'gate_proj.bias' in unpacked_state_dict and 'up_proj.bias' in unpacked_state_dict:
gate_proj_bias = unpacked_state_dict['gate_proj.bias']
up_proj_bias = unpacked_state_dict['up_proj.bias']
# Concatenate gate and up biases along dimension 0
packed_state_dict['swiglu.w12.bias'] = torch.cat([gate_proj_bias, up_proj_bias], dim=0)
# Copy down parameters as-is
if "down_proj.weight" in unpacked_state_dict:
packed_state_dict["swiglu.w3.weight"] = unpacked_state_dict['down_proj.weight']
if "down_proj.bias" in unpacked_state_dict:
packed_state_dict["swiglu.w3.bias"] = unpacked_state_dict['down_proj.bias']
for key in ['swiglu.w3.weight', 'swiglu.w3.bias']:
if key in unpacked_state_dict:
packed_state_dict[key] = unpacked_state_dict[key]
# Copy any other parameters that might exist
excluded_keys = [
'gate_proj.weight', 'gate_proj.bias',
'down_proj.weight', 'down_proj.bias',
'up_proj.weight', 'up_proj.bias',
'swiglu.w12.weight', 'swiglu.w12.bias',
'swiglu.w3.weight', 'swiglu.w3.bias',
]
for key, value in unpacked_state_dict.items():
if key not in excluded_keys:
packed_state_dict[key] = value
return packed_state_dict
def load_state_dict(self, state_dict, strict=True):
"""
Load state dict, handling both packed (w13) and unpacked (w1, w3) formats.
"""
# Check if this is an unpacked state dict (has w1 and w3 instead of w13)
has_unpacked_gate_up = 'gate_proj.weight' in state_dict and 'up_proj.weight' in state_dict
has_packed_swiglu = 'swiglu.w12.weight' in state_dict
if has_unpacked_gate_up and not has_packed_swiglu:
state_dict = self._convert_unpacked_to_packed_state_dict(state_dict)
return super().load_state_dict(state_dict, strict=strict)
def state_dict(self, destination=None, prefix='', keep_vars=False, packed=False):
"""
Return state dict in unpacked format by default for compatibility.
Set packed=True to get the internal packed format.
"""
if packed:
# Return the actual packed state dict
return super().state_dict(destination=destination, prefix=prefix, keep_vars=keep_vars)
else:
# Return unpacked format for compatibility
return self.get_unpacked_state_dict(destination=destination, prefix=prefix, keep_vars=keep_vars)
def get_unpacked_state_dict(self, destination=None, prefix='', keep_vars=False):
"""
Convert current packed state dict to unpacked format for compatibility.
"""
# Get the actual packed state dict first
packed_state_dict = super().state_dict(destination=None, prefix='', keep_vars=keep_vars)
if destination is None:
destination = OrderedDict()
# Handle w13 -> w1 and w3 conversion for weights
if f'{prefix}swiglu.w12.weight' in packed_state_dict:
w13_weight = packed_state_dict[f'{prefix}swiglu.w12.weight']
hidden_dim = w13_weight.shape[0] // 2
w1_weight, w3_weight = torch.split(w13_weight, hidden_dim, dim=0)
destination[f'{prefix}gate_proj.weight'] = w1_weight if not keep_vars else w1_weight.detach().requires_grad_(w1_weight.requires_grad)
destination[f'{prefix}up_proj.weight'] = w3_weight if not keep_vars else w3_weight.detach().requires_grad_(w3_weight.requires_grad)
# Handle w13 -> w1 and w3 conversion for biases (if they exist)
if f'{prefix}swiglu.w12.bias' in packed_state_dict:
w13_bias = packed_state_dict[f'{prefix}swiglu.w12.bias']
hidden_dim = w13_bias.shape[0] // 2
w1_bias, w3_bias = torch.split(w13_bias, hidden_dim, dim=0)
destination[f'{prefix}gate_proj.bias'] = w1_bias if not keep_vars else w1_bias.detach().requires_grad_(w1_bias.requires_grad)
destination[f'{prefix}up_proj.bias'] = w3_bias if not keep_vars else w3_bias.detach().requires_grad_(w3_bias.requires_grad)
# Copy w2 parameters as-is
for param_name in ['weight', 'bias']:
key = f'{prefix}swiglu.w3.{param_name}'
if key in packed_state_dict:
destination[f'{prefix}down_proj.{param_name}'] = packed_state_dict[key]
# Copy any other parameters
excluded_prefixes = [f'{prefix}swiglu.w12.', f'{prefix}swiglu.w3.']
for key, value in packed_state_dict.items():
if not any(key.startswith(excluded_prefix) for excluded_prefix in excluded_prefixes) and key not in destination:
destination[key] = value
return destination
def forward(self, x):
return self.swiglu(x)