Files
VLA/wall_x/fusions/ops.py
T
Starrick Liu 421db17d53 feat: Major optimization and robustness improvements (#31)
This release introduces significant performance optimizations, memory efficiency
improvements, and enhanced system robustness:

🚀 Performance Optimizations:
- Add three new fused CUDA kernels (rope_index, rot_pos_emb, get_window_index)
  for accelerated multimodal preprocessing
- Implement FSDP2 support for distributed training with improved memory efficiency
- Add Torch.compile integration for additional performance gains
- Optimize memory usage: reduce peak allocation from 48GB to 24GB on 8-GPU setup

🔧 System Robustness:
- Fix missing token position inputs in prediction pipeline
- Add type-robust negation operations in RoPE CUDA kernels (half/bfloat16 support)
- Fix dataset root parameter initialization in LeRobot data loader
- Enhanced error handling and input validation across fusion operators

📚 Documentation & Usability:
- Add comprehensive memory usage benchmarks and hardware recommendations
- Update citation format with proper arXiv reference
- Improve training configuration documentation with quick start guide
- Add detailed API documentation for new fusion operators

🛠️ Technical Details:
- Version bump to 1.0.1
- New CUDA kernels: rope_index.cu, rot_pos.cu, window_index.cu
- FSDP2 state dict loading with distribute_tensor support
- Enhanced multimodal RoPE with 3D position encoding
- Window attention optimization for Vision Transformers

Breaking Changes: None - all changes are backward compatible
2025-09-17 23:09:20 +08:00

743 lines
26 KiB
Python

import torch
import warnings
from wall_x.fusions import backend
class AsymmetricDualExpertGemm(torch.autograd.Function):
@staticmethod
def forward(
ctx, input_expert0, input_expert1, weight_expert0, weight_expert1, trans_b=False
):
"""
Forward pass for asymmetric dual expert GEMM.
Args:
input_expert0: Expert 0 input [m0, k]
input_expert1: Expert 1 input [m1, k]
weight_expert0: Expert 0 weight [k, n0] or [n0, k] if trans_b=True
weight_expert1: Expert 1 weight [k, n1] or [n1, k] if trans_b=True
trans_b: Whether to transpose the weight matrices
Returns:
Tuple of (output_expert0, output_expert1)
"""
# Validate inputs
assert input_expert0.dim() == 2, "input_expert0 must be 2D"
assert input_expert1.dim() == 2, "input_expert1 must be 2D"
assert weight_expert0.dim() == 2, "weight_expert0 must be 2D"
assert weight_expert1.dim() == 2, "weight_expert1 must be 2D"
# Dimension validation depends on trans_b
if trans_b:
assert input_expert0.size(1) == weight_expert0.size(
1
), "Expert 0 dimension mismatch (trans_b=True)"
assert input_expert1.size(1) == weight_expert1.size(
1
), "Expert 1 dimension mismatch (trans_b=True)"
else:
assert input_expert0.size(1) == weight_expert0.size(
0
), "Expert 0 dimension mismatch (trans_b=False)"
assert input_expert1.size(1) == weight_expert1.size(
0
), "Expert 1 dimension mismatch (trans_b=False)"
# Save tensors and trans_b for backward pass
ctx.save_for_backward(
input_expert0, input_expert1, weight_expert0, weight_expert1
)
ctx.trans_b = trans_b
# Allocate output tensors
m0 = input_expert0.size(0)
m1 = input_expert1.size(0)
n0 = weight_expert0.size(0) if trans_b else weight_expert0.size(1)
n1 = weight_expert1.size(0) if trans_b else weight_expert1.size(1)
output_expert0 = torch.empty(
m0, n0, device=input_expert0.device, dtype=input_expert0.dtype
)
output_expert1 = torch.empty(
m1, n1, device=input_expert1.device, dtype=input_expert1.dtype
)
# Call the backend C++ function
backend.asym_dual_gmm_separated(
input_expert0,
input_expert1,
weight_expert0,
weight_expert1,
output_expert0,
output_expert1,
trans_b=trans_b,
)
return output_expert0, output_expert1
@staticmethod
def backward(ctx, grad_output_expert0, grad_output_expert1):
"""
Optimized backward pass using specialized kernels.
Always computes all gradients to minimize kernel calls.
"""
grad_output_expert0 = grad_output_expert0.contiguous()
grad_output_expert1 = grad_output_expert1.contiguous()
input_expert0, input_expert1, weight_expert0, weight_expert1 = ctx.saved_tensors
trans_b = ctx.trans_b
# Always allocate all gradient tensors (no conditional computation)
grad_input_expert0 = torch.empty_like(input_expert0)
grad_input_expert1 = torch.empty_like(input_expert1)
grad_weight_expert0 = torch.empty_like(weight_expert0)
grad_weight_expert1 = torch.empty_like(weight_expert1)
# Compute input gradients: grad_input = grad_output @ weight^T (if trans_b=False)
# = grad_output @ weight (if trans_b=True)
backend.asym_dual_gmm_separated(
grad_output_expert0,
grad_output_expert1,
weight_expert0,
weight_expert1,
grad_input_expert0,
grad_input_expert1,
trans_a=False,
trans_b=not trans_b,
)
# Compute weight gradients
if trans_b:
# When trans_b=True in forward: output = input @ weight^T
# So grad_weight^T = input^T @ grad_output
# Which means grad_weight = grad_output^T @ input
backend.asym_dual_gmm_separated(
grad_output_expert0,
grad_output_expert1,
input_expert0,
input_expert1,
grad_weight_expert0,
grad_weight_expert1,
trans_a=True,
trans_b=False,
)
else:
# When trans_b=False in forward: output = input @ weight
# So grad_weight = input^T @ grad_output
backend.asym_dual_gmm_separated(
input_expert0,
input_expert1,
grad_output_expert0,
grad_output_expert1,
grad_weight_expert0,
grad_weight_expert1,
trans_a=True,
trans_b=False,
)
return (
grad_input_expert0,
grad_input_expert1,
grad_weight_expert0,
grad_weight_expert1,
None,
)
def asym_dual_gmm(
input_expert0, input_expert1, weight_expert0, weight_expert1, trans_b=False
):
"""
Convenience function for asymmetric dual expert GEMM.
Args:
input_expert0: Expert 0 input [m0, k]
input_expert1: Expert 1 input [m1, k]
weight_expert0: Expert 0 weight [k, n0] or [n0, k] if trans_b=True
weight_expert1: Expert 1 weight [k, n1] or [n1, k] if trans_b=True
trans_b: Whether to transpose the weight matrices
Returns:
Tuple of (output_expert0, output_expert1)
"""
return AsymmetricDualExpertGemm.apply(
input_expert0, input_expert1, weight_expert0, weight_expert1, trans_b
)
################################################################################################
##
## PermuteMoE topK
##
################################################################################################
class PermuteMoE_topK(torch.autograd.Function):
workspace_fw = None
dtype = None
max_expanded_token_num = 0
@staticmethod
def forward(
ctx,
input_act: torch.Tensor,
indices: torch.Tensor,
num_out_tokens: int,
max_token_num: int,
):
"""
indices: for topK=1, indices in a 1-d tensor of shape [num_tokens],
otherwise, it's a 2-d tensor of shape [num_tokens, topK]
"""
# Empty input check
if not input_act.numel():
return input_act, None
# For top1 case, view the indices as 2D tensor to unify the shape for topk>=2 cases.
if indices.dim() == 1:
indices = indices.view(-1, 1)
# Device check
if input_act.is_cpu:
raise RuntimeError(
"[Error] The input `input_act` of permute_topK op is on the device: CPU!"
)
if indices.is_cpu:
warnings.warn(
"The input `indices` of permute_topK op is on the device: CPU!"
)
# Shape check
if input_act.size(0) != indices.size(0):
raise RuntimeError(
f"[Error] permute_topK op input `indices` shape mismatch! "
f"Expect {input_act.size(0)}, but got {indices.size(0)}."
)
# Data type check
if indices.dtype != torch.int32:
warnings.warn(
f"The data type of the input `indices` of permute_topK op is {indices.dtype}! "
"The recommended type is torch.int32."
)
indices = indices.to(torch.int32)
# Contiguous check
if not input_act.is_contiguous():
warnings.warn("The input `input_act` of permute_topK op is discontiguous!")
input_act = input_act.contiguous()
if not indices.is_contiguous():
warnings.warn("The input `indices` of permute_topK op is discontiguous!")
indices = indices.contiguous()
num_topK = indices.size(1)
input_max_expanded_token_num = max(max_token_num, input_act.size(0)) * num_topK
if PermuteMoE_topK.max_expanded_token_num < input_max_expanded_token_num:
PermuteMoE_topK.max_expanded_token_num = input_max_expanded_token_num
PermuteMoE_topK.workspace_fw = []
if PermuteMoE_topK.dtype != input_act.dtype:
PermuteMoE_topK.dtype = input_act.dtype
PermuteMoE_topK.workspace_fw = []
permuted_act, row_id_map, PermuteMoE_topK.workspace_fw = backend.permute(
input_act,
indices,
num_out_tokens,
PermuteMoE_topK.workspace_fw,
PermuteMoE_topK.max_expanded_token_num,
)
ctx.row_id_map = row_id_map
ctx.num_tokens = indices.size(0)
ctx.num_topK = num_topK
return permuted_act, row_id_map
@staticmethod
def backward(ctx, permuted_act_grad, _):
# Empty input check
if not permuted_act_grad.numel():
return permuted_act_grad, None, None, None
if not permuted_act_grad.is_contiguous():
permuted_act_grad = permuted_act_grad.contiguous()
row_id_map = ctx.row_id_map
num_tokens = ctx.num_tokens
num_topK = ctx.num_topK
unpermuted_act_grad = backend.unpermute(
permuted_act_grad, row_id_map, torch.tensor([]), num_tokens, num_topK
)
return unpermuted_act_grad, None, None, None
################################################################################################
##
## UnpermuteMoE topK
##
################################################################################################
class UnpermuteMoE_topK(torch.autograd.Function):
@staticmethod
def forward(
ctx,
input_act: torch.Tensor,
row_id_map: torch.Tensor,
probs: torch.Tensor = None,
):
# Empty input check
if not input_act.numel():
ctx.probs = probs
return input_act
# Device check
if input_act.is_cpu:
raise RuntimeError(
"[Error] The input `input_act` of unpermute_topK op is on the device: CPU!"
)
if row_id_map.is_cpu:
warnings.warn(
"The input `row_id_map` of unpermute_topK op is on the device: CPU!"
)
row_id_map = row_id_map.cuda()
if probs is not None and probs.is_cpu:
warnings.warn(
"The input `probs` of unpermute_topK op is on the device: CPU!"
)
probs = probs.cuda()
# Shape check
if probs is not None and row_id_map.size(0) != probs.size(0) * probs.size(1):
raise RuntimeError(
f"[Error] unpermute_topK op input `probs` shape mismatch! "
f"Expect {row_id_map.size(0)}, but got {probs.size(0) * probs.size(1)}."
)
# Data type check
if row_id_map.dtype != torch.int32:
warnings.warn(
f"The data type of the input `row_id_map` of unpermute_topK op is {row_id_map.dtype}! "
"The recommended type is torch.int32."
)
row_id_map = row_id_map.to(torch.int32)
if probs is not None and probs.dtype != torch.float32:
warnings.warn(
f"The data type of the input `probs` of unpermute_topK op is {probs.dtype}! "
"The recommended type is torch.float32."
)
probs = probs.to(torch.float32)
# Contiguous check
if not input_act.is_contiguous():
warnings.warn(
"The input `input_act` of unpermute_topK op is discontiguous!"
)
input_act = input_act.contiguous()
if not row_id_map.is_contiguous():
warnings.warn(
"The input `row_id_map` of unpermute_topK op is discontiguous!"
)
row_id_map = row_id_map.contiguous()
if probs is not None and not probs.is_contiguous():
warnings.warn("The input `probs` of unpermute_topK op is discontiguous!")
probs = probs.contiguous()
num_tokens = probs.size(0) if probs is not None else input_act.size(0)
num_topK = probs.size(1) if probs is not None else 1
unpermuted_output = backend.unpermute(
input_act,
row_id_map,
probs if probs is not None else torch.tensor([]),
num_tokens,
num_topK,
)
ctx.save_for_backward(input_act, row_id_map, probs)
return unpermuted_output
@staticmethod
def backward(ctx, unpermuted_act_grad):
# Empty input check
if not unpermuted_act_grad.numel():
return unpermuted_act_grad, None, ctx.probs
if not unpermuted_act_grad.is_contiguous():
unpermuted_act_grad = unpermuted_act_grad.contiguous()
input_act, row_id_map, probs = ctx.saved_tensors
act_grad = None
if ctx.needs_input_grad[0]:
act_grad, prob_grad = backend.unpermute_bwd(
unpermuted_act_grad, input_act, row_id_map, probs
)
if not ctx.needs_input_grad[2]:
prob_grad = None
return act_grad, None, prob_grad
def permute(input_act, indices, num_out_tokens=None, max_token_num=0):
num_out_tokens = 0 if num_out_tokens is None else num_out_tokens
return PermuteMoE_topK.apply(input_act, indices, num_out_tokens, max_token_num)
def unpermute(input_act, row_id_map, probs=None):
return UnpermuteMoE_topK.apply(input_act, row_id_map, probs)
################################################################################################
##
## mutlimodal RoPE
##
################################################################################################
class MultimodalRoPE(torch.autograd.Function):
@staticmethod
def forward(
ctx,
q: torch.Tensor,
k: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mrope_section: list,
):
# Device check
if q.is_cpu:
raise RuntimeError(
"[Error] The input `q` of multimodal_rope op is on the device: CPU!"
)
if k.is_cpu:
raise RuntimeError(
"[Error] The input `k` of multimodal_rope op is on the device: CPU!"
)
if cos.is_cpu:
raise RuntimeError(
"[Error] The input `cos` of multimodal_rope op is on the device: CPU!"
)
if sin.is_cpu:
raise RuntimeError(
"[Error] The input `sin` of multimodal_rope op is on the device: CPU!"
)
if len(mrope_section) != 3:
raise RuntimeError(
"[Error] The input `mrope_section` of multimodal_rope op must be a list of 3 integers!"
)
# Contiguous check
if not q.is_contiguous():
warnings.warn("The input `q` of multimodal_rope op is discontiguous!")
q = q.contiguous()
if not k.is_contiguous():
warnings.warn("The input `k` of multimodal_rope op is discontiguous!")
k = k.contiguous()
if not cos.is_contiguous():
warnings.warn("The input `cos` of multimodal_rope op is discontiguous!")
cos = cos.contiguous()
if not sin.is_contiguous():
warnings.warn("The input `sin` of multimodal_rope op is discontiguous!")
sin = sin.contiguous()
# Prepare mrope_section_doubled
mrope_section_doubled = [x * 2 for x in mrope_section]
# Create output tensors
q_out = torch.empty_like(q)
k_out = torch.empty_like(k)
backend.rope(q, k, cos, sin, q_out, k_out, mrope_section_doubled)
ctx.save_for_backward(q, k, cos, sin)
ctx.mrope_section_doubled = mrope_section_doubled
return q_out, k_out
@staticmethod
def backward(ctx, grad_q_out, grad_k_out):
if not grad_q_out.is_contiguous():
grad_q_out = grad_q_out.contiguous()
if not grad_k_out.is_contiguous():
grad_k_out = grad_k_out.contiguous()
q, k, cos, sin = ctx.saved_tensors
grad_q = None
grad_k = None
if ctx.needs_input_grad[0]:
grad_q = torch.empty_like(q)
if ctx.needs_input_grad[1]:
grad_k = torch.empty_like(k)
if grad_q is not None or grad_k is not None:
backend.rope_bwd(
grad_q_out,
grad_k_out,
q,
k,
cos,
sin,
grad_q if grad_q is not None else torch.empty_like(q),
grad_k if grad_k is not None else torch.empty_like(k),
ctx.mrope_section_doubled,
)
return grad_q, grad_k, None, None, None
def multimodal_rope(q, k, cos, sin, mrope_section):
return MultimodalRoPE.apply(q, k, cos, sin, mrope_section)
################################################################################################
##
## RoPE Index 3D
##
################################################################################################
def get_rope_index(
input_ids: torch.Tensor,
spatial_merge_size: int,
image_token_id: int,
video_token_id: int,
vision_start_token_id: int,
tokens_per_second: float,
image_grid_thw: torch.Tensor = None,
video_grid_thw: torch.Tensor = None,
second_per_grid_ts: torch.Tensor = None,
attention_mask: torch.Tensor = None,
):
"""
Generate 3D RoPE position indices for multimodal transformer inputs.
Computes position indices for text, image, and video tokens to enable proper
spatial-temporal position encoding in multimodal transformers with RoPE.
Args:
input_ids (torch.Tensor): Input token sequence of shape [batch_size, seq_len]
Must be LongTensor on CUDA device
spatial_merge_size (int): Spatial merge size for patch grouping (must be positive)
image_token_id (int): Token ID representing image patches
video_token_id (int): Token ID representing video frames
vision_start_token_id (int): Token ID marking start of vision sequences
tokens_per_second (float): Temporal scaling factor for video sequences (must be positive)
image_grid_thw (torch.Tensor, optional): Image grid dimensions of shape [num_images, 3] (T, H, W)
video_grid_thw (torch.Tensor, optional): Video grid dimensions of shape [num_videos, 3] (T, H, W)
second_per_grid_ts (torch.Tensor, optional): Video time intervals of shape [num_videos]
attention_mask (torch.Tensor, optional): Attention mask of shape [batch_size, seq_len]
Returns:
Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
- position_ids: 3D position indices of shape [3, batch_size, seq_len]
- mrope_position_deltas: mRoPE position deltas of shape [batch_size, 1]
Raises:
TypeError: If input_ids is not a torch.Tensor
ValueError: If input dimensions are incorrect or tensors not on CUDA
"""
# Input validation
if not isinstance(input_ids, torch.Tensor):
raise TypeError("input_ids must be a torch.Tensor")
if input_ids.dim() != 2:
raise ValueError("input_ids must be 2D tensor (batch_size, seq_len)")
if not input_ids.is_cuda:
raise ValueError("input_ids must be on CUDA device")
# Parameter validation
if not isinstance(spatial_merge_size, int) or spatial_merge_size <= 0:
raise ValueError(
f"spatial_merge_size must be positive integer, got {spatial_merge_size}"
)
if not isinstance(tokens_per_second, (int, float)) or tokens_per_second <= 0:
raise ValueError(
f"tokens_per_second must be positive number, got {tokens_per_second}"
)
return backend.get_rope_index(
input_ids,
image_grid_thw,
video_grid_thw,
second_per_grid_ts,
attention_mask,
spatial_merge_size,
image_token_id,
video_token_id,
vision_start_token_id,
float(tokens_per_second),
)
################################################################################################
##
## Fused Rotary Position Embedding
##
################################################################################################
def rot_pos_emb(
inv_freq: torch.Tensor,
grid_thw: torch.Tensor,
spatial_merge_size: int,
) -> torch.Tensor:
"""
Compute fused rotary position embeddings using optimized CUDA kernel.
This function fuses all rotary position embedding computations into a single
CUDA kernel for improved performance with spatial-temporal grids.
Args:
inv_freq (torch.Tensor): Inverse frequencies tensor of shape [dim/2]
Contains precomputed 1.0 / (theta ** (torch.arange(0, dim, 2) / dim))
Must be float32 on CUDA device
grid_thw (torch.Tensor): Grid dimensions tensor of shape [num_grids, 3]
Each row contains (T, H, W) for temporal, height, width dimensions
Supports int32 or int64 on CUDA device
spatial_merge_size (int): Spatial merge size for token grouping (must be positive)
Returns:
torch.Tensor: Rotary position embeddings of shape [total_tokens, dim]
where dim = 2 * len(inv_freq)
First half contains h_pos frequencies, second half contains w_pos frequencies
Raises:
TypeError: If inputs are not torch.Tensor or spatial_merge_size not int
ValueError: If tensor dimensions incorrect, not on CUDA, or devices mismatch
RuntimeError: If CUDA kernel execution fails
"""
# Type checking
if not isinstance(inv_freq, torch.Tensor):
raise TypeError(f"inv_freq must be a torch.Tensor, got {type(inv_freq)}")
if not isinstance(grid_thw, torch.Tensor):
raise TypeError(f"grid_thw must be a torch.Tensor, got {type(grid_thw)}")
# Dimension checking
if inv_freq.dim() != 1:
raise ValueError(
f"inv_freq must be 1-dimensional, got {inv_freq.dim()}D tensor"
)
if grid_thw.dim() != 2:
raise ValueError(
f"grid_thw must be 2-dimensional, got {grid_thw.dim()}D tensor"
)
if grid_thw.size(1) != 3:
raise ValueError(
f"grid_thw must have shape [num_grids, 3], got shape {list(grid_thw.shape)}"
)
# Device checking
if not inv_freq.is_cuda:
raise ValueError("inv_freq must be on CUDA device")
if not grid_thw.is_cuda:
raise ValueError("grid_thw must be on CUDA device")
# Ensure both tensors are on the same device
if inv_freq.device != grid_thw.device:
raise ValueError(
f"inv_freq and grid_thw must be on the same device, "
f"got {inv_freq.device} and {grid_thw.device}"
)
# Parameter validation
if not isinstance(spatial_merge_size, int):
raise TypeError(
f"spatial_merge_size must be an integer, got {type(spatial_merge_size)}"
)
if spatial_merge_size <= 0:
raise ValueError(
f"spatial_merge_size must be positive, got {spatial_merge_size}"
)
# Ensure inv_freq is float32 (the kernel expects float)
if inv_freq.dtype != torch.float32:
inv_freq = inv_freq.to(torch.float32)
# Call the CUDA backend
try:
return backend.rot_pos_emb(inv_freq, grid_thw, spatial_merge_size)
except RuntimeError as e:
raise RuntimeError(f"CUDA kernel execution failed: {str(e)}")
################################################################################################
##
## Fused Window Index Generation
##
################################################################################################
def get_window_index(
grid_thw: torch.Tensor,
window_size: int,
spatial_merge_size: int,
patch_size: int,
spatial_merge_unit: int = 1,
):
"""
Generate window attention indices for Vision Transformer architectures.
Computes window-based attention indices for hierarchical processing of vision
tokens, enabling efficient sliding window attention patterns in ViT models.
Args:
grid_thw (torch.Tensor): Grid specifications of shape [num_grids, 3] (T, H, W)
Must be or will be converted to int32 on CUDA device
window_size (int): Window size for attention computation
spatial_merge_size (int): Spatial merge size for patch grouping
patch_size (int): Size of vision patches in pixels
spatial_merge_unit (int, optional): Spatial merging unit size. Defaults to 1.
Returns:
Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
- window_index: Window indices tensor of shape [total_elements]
- cu_window_seqlens: Cumulative window sequence lengths of shape [num_windows + 1]
Raises:
AssertionError: If grid_thw dimensions are incorrect
Note:
Returns empty tensors if input grid is empty or no valid windows can be formed.
The function automatically converts input to CUDA int32 if needed.
"""
# Input validation
assert (
grid_thw.dim() == 2 and grid_thw.size(1) == 3
), f"grid_thw must have shape (num_grids, 3), got {grid_thw.shape}"
# Ensure input is on CUDA and int32 type
if not grid_thw.is_cuda:
grid_thw = grid_thw.cuda()
if grid_thw.dtype != torch.int32:
grid_thw = grid_thw.to(torch.int32)
# Calculate vit_merger_window_size
vit_merger_window_size = window_size // spatial_merge_size // patch_size
# Call CUDA backend
window_index, cu_window_seqlens = backend.get_window_index(
grid_thw,
spatial_merge_size,
vit_merger_window_size,
patch_size,
spatial_merge_unit,
)
return window_index, cu_window_seqlens