Kernels:

kernels-community
/

triton-moe

Trusted publisher

Kernel card Files Files and versions

xet

Community

triton-moe / tests /test_triton_moe.py

drbh

feat: fix backward op

399fb8c over 1 year ago

Raw History Blame Contribute Delete

35.4 kB

	import torch
	import torch.nn.functional as F
	import torch.nn as nn
	import time

	from triton_moe.layers import MoE


	# Copied from transformers.models.openai.modeling_openai
	class OpenaiExperts(nn.Module):
	def __init__(self, config):
	super().__init__()
	self.num_experts = config.num_local_experts
	self.intermediate_size = config.intermediate_size
	self.hidden_size = config.hidden_size
	self.expert_dim = self.intermediate_size
	self.gate_up_proj = nn.Parameter(
	torch.empty(self.num_experts, self.hidden_size, 2 * self.expert_dim)
	)
	self.gate_up_proj_bias = nn.Parameter(
	torch.empty(self.num_experts, 2 * self.expert_dim)
	)
	self.down_proj = nn.Parameter(
	torch.empty((self.num_experts, self.expert_dim, self.hidden_size))
	)
	self.down_proj_bias = nn.Parameter(
	torch.empty(self.num_experts, self.hidden_size)
	)
	self.alpha = 1.702

	def forward(
	self, hidden_states: torch.Tensor, router_indices=None, routing_weights=None
	) -> torch.Tensor:
	"""
	When training is is more efficient to just loop over the experts and compute the output for each expert
	as otherwise the memory would explode.

	For inference we can sacrifice some memory and compute the output for all experts at once. By repeating the inputs.

	Args:
	hidden_states (torch.Tensor): (batch_size * token_num, hidden_size)
	selected_experts (torch.Tensor): (batch_size * token_num, top_k)
	routing_weights (torch.Tensor): (batch_size * token_num, top_k)
	Returns:
	torch.Tensor
	"""
	if self.training:
	next_states = torch.zeros_like(
	hidden_states, dtype=hidden_states.dtype, device=hidden_states.device
	)
	with torch.no_grad():
	expert_mask = torch.nn.functional.one_hot(
	router_indices, num_classes=self.num_experts
	).permute(2, 1, 0)
	expert_hitted = torch.greater(
	expert_mask.sum(dim=(-1, -2)), 0
	).nonzero()
	for expert_idx in expert_hitted:
	with torch.no_grad():
	idx, top_x = torch.where(
	expert_mask[expert_idx][0]
	) # idx: top-1/top-2 indicator, top_x: token indices
	current_state = hidden_states[top_x] # (num_tokens, hidden_dim)
	gate_up = (
	current_state @ self.gate_up_proj[expert_idx]
	+ self.gate_up_proj_bias[expert_idx]
	) # (num_tokens, 2 * interm_dim)
	gate, up = gate_up.chunk(2, dim=-1) # (num_tokens, interm_dim)
	glu = gate * torch.sigmoid(
	gate * self.alpha
	) # (num_tokens, interm_dim)
	gated_output = (up + 1) * glu # (num_tokens, interm_dim)
	out = (
	gated_output @ self.down_proj[expert_idx]
	+ self.down_proj_bias[expert_idx]
	) # (num_tokens, hidden_dim)
	weighted_output = (
	out * routing_weights[top_x, idx, None]
	) # (num_tokens, hidden_dim)
	next_states.index_add_(
	0, top_x, weighted_output.to(hidden_states.dtype)[0]
	)
	else:
	hidden_states = hidden_states.repeat(self.num_experts, 1)
	hidden_states = hidden_states.view(self.num_experts, -1, self.hidden_size)
	gate_up = (
	torch.bmm(hidden_states, self.gate_up_proj)
	+ self.gate_up_proj_bias[..., None, :]
	)
	gate, up = gate_up.chunk(2, dim=-1) # not supported for DTensors
	glu = gate * torch.sigmoid(gate * self.alpha)
	next_states = (
	torch.bmm(((up + 1) * glu), self.down_proj)
	+ self.down_proj_bias[..., None, :]
	)
	next_states = next_states.view(-1, self.hidden_size)
	return next_states


	def test_moe_forward():

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	requires_grad=True,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True) # Normalize

	# Initialize parameters
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	ref_layer.gate_up_proj = gate_up_proj
	ref_layer.gate_up_proj_bias = gate_up_proj_bias
	ref_layer.down_proj = down_proj
	ref_layer.down_proj_bias = down_proj_bias

	with torch.no_grad():
	old_output = ref_layer(hidden_states, router_idx, router_wt)
	output = layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)

	assert old_output.shape == output.shape, "Output shapes do not match"

	diff = (old_output - output).abs()
	avg_diff = diff.mean()

	print(f"Average difference: {avg_diff.item()}")
	# Average difference: 0.009219333529472351

	print(f"Max difference: {diff.max().item()}")
	# Max difference: 0.09375

	# TODO: Improve the precision
	assert torch.allclose(
	# old_output, output, rtol=1e-3, atol=1e-3
	old_output,
	output,
	rtol=1e-1,
	atol=1e-1,
	), "Outputs do not match between the two implementations"


	def test_moe_backward_grad():
	"""Simple backward test comparing gradients between MoE and OpenaiExperts."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	requires_grad=True,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True) # Normalize

	# Initialize parameters (shared between both implementations)
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()

	output = layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)

	# Backward pass
	loss = output.sum()

	loss.backward()

	# Check that gradients exist
	print(f"gate_up_proj.grad exists: {gate_up_proj.grad is not None}")
	print(f"gate_up_proj_bias.grad exists: {gate_up_proj_bias.grad is not None}")
	print(f"down_proj.grad exists: {down_proj.grad is not None}")
	print(f"down_proj_bias.grad exists: {down_proj_bias.grad is not None}")

	# Simple check: if gradients exist, the backward pass worked
	assert (
	gate_up_proj.grad is not None
	), "gate_up_proj gradient is None - custom MoE not using this parameter"
	assert (
	gate_up_proj_bias.grad is not None
	), "gate_up_proj_bias gradient is None - custom MoE not using this parameter"
	assert down_proj.grad is not None, "down_proj gradient is None"
	assert down_proj_bias.grad is not None, "down_proj_bias gradient is None"


	def test_moe_backward():
	"""Simple backward test comparing gradients between MoE and OpenaiExperts."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	requires_grad=True,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True) # Normalize

	# Initialize parameters (shared between both implementations)
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	# Set reference layer parameters and clone to avoid mutating the original parameters
	ref_layer.gate_up_proj = nn.Parameter(
	gate_up_proj.clone().detach().requires_grad_(True)
	)
	ref_layer.gate_up_proj_bias = nn.Parameter(
	gate_up_proj_bias.clone().detach().requires_grad_(True)
	)
	ref_layer.down_proj = nn.Parameter(down_proj.clone().detach().requires_grad_(True))
	ref_layer.down_proj_bias = nn.Parameter(
	down_proj_bias.clone().detach().requires_grad_(True)
	)

	# Forward pass
	ref_output = ref_layer(hidden_states, router_idx, router_wt)
	output = layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)

	# Backward pass
	ref_loss = ref_output.sum()
	loss = output.sum()

	ref_loss.backward(retain_graph=True)
	loss.backward()

	# Check that gradients exist
	print(f"gate_up_proj.grad exists: {gate_up_proj.grad is not None}")
	print(f"gate_up_proj_bias.grad exists: {gate_up_proj_bias.grad is not None}")
	print(f"down_proj.grad exists: {down_proj.grad is not None}")
	print(f"down_proj_bias.grad exists: {down_proj_bias.grad is not None}")
	print(f"hidden_states.grad exists: {hidden_states.grad is not None}")

	# Simple check: if gradients exist, the backward pass worked
	assert (
	gate_up_proj.grad is not None
	), "gate_up_proj gradient is None - custom MoE not using this parameter"
	assert (
	gate_up_proj_bias.grad is not None
	), "gate_up_proj_bias gradient is None - custom MoE not using this parameter"
	assert down_proj.grad is not None, "down_proj gradient is None"
	assert down_proj_bias.grad is not None, "down_proj_bias gradient is None"
	assert hidden_states.grad is not None, "hidden_states gradient is None"

	print("✓ Backward test passed - all parameters have gradients")

	# Sanity check on gate up gradients
	print("10 elements from gate_up_proj gradients:")
	print(gate_up_proj.grad.flatten()[:10])
	print("10 elements from ref_layer.gate_up_proj gradients:")
	print(ref_layer.gate_up_proj.grad.flatten()[:10])

	# Compare the values and ensure they are close enough
	assert torch.allclose(
	ref_layer.gate_up_proj.grad,
	gate_up_proj.grad,
	rtol=1e-1,
	atol=1e-1,
	), "gate_up_proj gradients do not match between implementations"
	assert torch.allclose(
	ref_layer.gate_up_proj_bias.grad,
	gate_up_proj_bias.grad,
	rtol=1e-1,
	atol=1e-1,
	), "gate_up_proj_bias gradients do not match between implementations"
	assert torch.allclose(
	ref_layer.down_proj.grad,
	down_proj.grad,
	rtol=1e-1,
	atol=1e-1,
	), "down_proj gradients do not match between implementations"
	assert torch.allclose(
	ref_layer.down_proj_bias.grad,
	down_proj_bias.grad,
	rtol=1e-1,
	atol=1e-1,
	), "down_proj_bias gradients do not match between implementations"


	def test_moe_backward_benchmark():
	"""Benchmark backward pass performance between MoE and OpenaiExperts."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2
	num_warmup = 5
	num_runs = 20

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	requires_grad=True,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True) # Normalize

	# Initialize parameters
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	# Set reference layer parameters
	ref_layer.gate_up_proj = gate_up_proj
	ref_layer.gate_up_proj_bias = gate_up_proj_bias
	ref_layer.down_proj = down_proj
	ref_layer.down_proj_bias = down_proj_bias

	def benchmark_ref_backward():
	"""Benchmark reference implementation backward pass."""
	hidden_states_copy = hidden_states.clone().detach().requires_grad_(True)

	# Clear gradients
	for param in [gate_up_proj, gate_up_proj_bias, down_proj, down_proj_bias]:
	if param.grad is not None:
	param.grad.zero_()

	# Forward + Backward
	output = ref_layer(hidden_states_copy, router_idx, router_wt)
	loss = output.sum()
	loss.backward()

	return loss.item()

	def benchmark_custom_backward():
	"""Benchmark custom implementation backward pass."""
	hidden_states_copy = hidden_states.clone().detach().requires_grad_(True)

	# Clear gradients
	for param in [gate_up_proj, gate_up_proj_bias, down_proj, down_proj_bias]:
	if param.grad is not None:
	param.grad.zero_()

	# Forward + Backward
	output = layer(
	hidden_states_copy,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)
	loss = output.sum()
	loss.backward()

	return loss.item()

	# Warmup
	print("Warming up...")
	for _ in range(num_warmup):
	benchmark_ref_backward()
	benchmark_custom_backward()

	# Clear cache and synchronize
	torch.cuda.empty_cache()
	torch.cuda.synchronize()

	# Benchmark reference implementation
	print(f"Benchmarking reference implementation ({num_runs} runs)...")
	ref_times = []

	for i in range(num_runs):
	torch.cuda.synchronize()
	start_time = time.perf_counter()

	benchmark_ref_backward()

	torch.cuda.synchronize()
	end_time = time.perf_counter()

	ref_times.append((end_time - start_time) * 1000) # Convert to ms

	if (i + 1) % 5 == 0:
	print(f" Completed {i + 1}/{num_runs} runs")

	# Clear cache and synchronize
	torch.cuda.empty_cache()
	torch.cuda.synchronize()

	# Benchmark custom implementation
	print(f"Benchmarking custom implementation ({num_runs} runs)...")
	custom_times = []

	for i in range(num_runs):
	torch.cuda.synchronize()
	start_time = time.perf_counter()

	benchmark_custom_backward()

	torch.cuda.synchronize()
	end_time = time.perf_counter()

	custom_times.append((end_time - start_time) * 1000) # Convert to ms

	if (i + 1) % 5 == 0:
	print(f" Completed {i + 1}/{num_runs} runs")

	# Calculate statistics
	ref_mean = sum(ref_times) / len(ref_times)
	ref_std = (sum((t - ref_mean) 2 for t in ref_times) / len(ref_times)) 0.5
	ref_min = min(ref_times)
	ref_max = max(ref_times)

	custom_mean = sum(custom_times) / len(custom_times)
	custom_std = (
	sum((t - custom_mean) ** 2 for t in custom_times) / len(custom_times)
	) ** 0.5
	custom_min = min(custom_times)
	custom_max = max(custom_times)

	speedup = ref_mean / custom_mean

	# Print results
	print("\n" + "=" * 80)
	print("BACKWARD PASS BENCHMARK RESULTS")
	print("=" * 80)
	print(f"Configuration:")
	print(f" - Experts: {num_experts}")
	print(f" - Hidden size: {hidden_size}")
	print(f" - Expert dim: {expert_dim}")
	print(f" - Batch tokens: {batch_tokens}")
	print(f" - Top-k: {topk}")
	print(f" - Runs: {num_runs}")
	print()

	print(f"Reference Implementation (OpenaiExperts):")
	print(f" - Mean: {ref_mean:.3f} ms")
	print(f" - Std: {ref_std:.3f} ms")
	print(f" - Min: {ref_min:.3f} ms")
	print(f" - Max: {ref_max:.3f} ms")
	print()

	print(f"Custom Implementation (MoE):")
	print(f" - Mean: {custom_mean:.3f} ms")
	print(f" - Std: {custom_std:.3f} ms")
	print(f" - Min: {custom_min:.3f} ms")
	print(f" - Max: {custom_max:.3f} ms")
	print()

	print(f"Speedup: {speedup:.2f}x")

	if speedup > 1.0:
	print(f"✓ Custom implementation is {speedup:.2f}x faster")
	else:
	print(f"✗ Custom implementation is {1/speedup:.2f}x slower")

	print("=" * 80)

	# Optional: Save detailed timing data
	print(f"\nDetailed timings (ms):")
	print(f"Reference: {ref_times}")
	print(f"Custom: {custom_times}")


	def test_moe_backward_benchmark_memory():
	"""Benchmark memory usage during backward pass."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	requires_grad=True,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True)

	# Initialize parameters
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	# Set reference layer parameters
	ref_layer.gate_up_proj = gate_up_proj
	ref_layer.gate_up_proj_bias = gate_up_proj_bias
	ref_layer.down_proj = down_proj
	ref_layer.down_proj_bias = down_proj_bias

	# Measure memory for reference implementation
	torch.cuda.empty_cache()
	torch.cuda.reset_peak_memory_stats()

	hidden_states_copy = hidden_states.clone().detach().requires_grad_(True)
	output = ref_layer(hidden_states_copy, router_idx, router_wt)
	loss = output.sum()
	loss.backward()

	ref_memory = torch.cuda.max_memory_allocated() / 1024**3 # GB

	# Measure memory for custom implementation
	torch.cuda.empty_cache()
	torch.cuda.reset_peak_memory_stats()

	hidden_states_copy = hidden_states.clone().detach().requires_grad_(True)
	output = layer(
	hidden_states_copy,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)
	loss = output.sum()
	loss.backward()

	custom_memory = torch.cuda.max_memory_allocated() / 1024**3 # GB

	print("\n" + "=" * 60)
	print("MEMORY USAGE BENCHMARK")
	print("=" * 60)
	print(f"Reference implementation: {ref_memory:.3f} GB")
	print(f"Custom implementation: {custom_memory:.3f} GB")
	print(f"Memory ratio: {custom_memory/ref_memory:.3f}x")

	if custom_memory < ref_memory:
	print(f"✓ Custom uses {(1 - custom_memory/ref_memory)*100:.1f}% less memory")
	else:
	print(f"✗ Custom uses {(custom_memory/ref_memory - 1)*100:.1f}% more memory")
	print("=" * 60)


	#################


	def test_moe_forward_benchmark():
	"""Benchmark forward pass performance between MoE and OpenaiExperts."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2
	num_warmup = 5
	num_runs = 50

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True) # Normalize

	# Initialize parameters
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	# Set reference layer parameters
	ref_layer.gate_up_proj = gate_up_proj
	ref_layer.gate_up_proj_bias = gate_up_proj_bias
	ref_layer.down_proj = down_proj
	ref_layer.down_proj_bias = down_proj_bias

	def benchmark_ref_forward():
	"""Benchmark reference implementation forward pass."""
	with torch.no_grad():
	output = ref_layer(hidden_states, router_idx, router_wt)
	return output

	def benchmark_custom_forward():
	"""Benchmark custom implementation forward pass."""
	with torch.no_grad():
	output = layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)
	return output

	# Warmup
	print("Warming up...")
	for _ in range(num_warmup):
	benchmark_ref_forward()
	benchmark_custom_forward()

	# Clear cache and synchronize
	torch.cuda.empty_cache()
	torch.cuda.synchronize()

	# Benchmark reference implementation
	print(f"Benchmarking reference implementation ({num_runs} runs)...")
	ref_times = []

	for i in range(num_runs):
	torch.cuda.synchronize()
	start_time = time.perf_counter()

	benchmark_ref_forward()

	torch.cuda.synchronize()
	end_time = time.perf_counter()

	ref_times.append((end_time - start_time) * 1000) # Convert to ms

	if (i + 1) % 10 == 0:
	print(f" Completed {i + 1}/{num_runs} runs")

	# Clear cache and synchronize
	torch.cuda.empty_cache()
	torch.cuda.synchronize()

	# Benchmark custom implementation
	print(f"Benchmarking custom implementation ({num_runs} runs)...")
	custom_times = []

	for i in range(num_runs):
	torch.cuda.synchronize()
	start_time = time.perf_counter()

	benchmark_custom_forward()

	torch.cuda.synchronize()
	end_time = time.perf_counter()

	custom_times.append((end_time - start_time) * 1000) # Convert to ms

	if (i + 1) % 10 == 0:
	print(f" Completed {i + 1}/{num_runs} runs")

	# Calculate statistics
	ref_mean = sum(ref_times) / len(ref_times)
	ref_std = (sum((t - ref_mean) 2 for t in ref_times) / len(ref_times)) 0.5
	ref_min = min(ref_times)
	ref_max = max(ref_times)

	custom_mean = sum(custom_times) / len(custom_times)
	custom_std = (
	sum((t - custom_mean) ** 2 for t in custom_times) / len(custom_times)
	) ** 0.5
	custom_min = min(custom_times)
	custom_max = max(custom_times)

	speedup = ref_mean / custom_mean

	# Print results
	print("\n" + "=" * 80)
	print("FORWARD PASS BENCHMARK RESULTS")
	print("=" * 80)
	print(f"Configuration:")
	print(f" - Experts: {num_experts}")
	print(f" - Hidden size: {hidden_size}")
	print(f" - Expert dim: {expert_dim}")
	print(f" - Batch tokens: {batch_tokens}")
	print(f" - Top-k: {topk}")
	print(f" - Runs: {num_runs}")
	print()

	print(f"Reference Implementation (OpenaiExperts):")
	print(f" - Mean: {ref_mean:.3f} ms")
	print(f" - Std: {ref_std:.3f} ms")
	print(f" - Min: {ref_min:.3f} ms")
	print(f" - Max: {ref_max:.3f} ms")
	print()

	print(f"Custom Implementation (MoE):")
	print(f" - Mean: {custom_mean:.3f} ms")
	print(f" - Std: {custom_std:.3f} ms")
	print(f" - Min: {custom_min:.3f} ms")
	print(f" - Max: {custom_max:.3f} ms")
	print()

	print(f"Speedup: {speedup:.2f}x")

	if speedup > 1.0:
	print(f"✓ Custom implementation is {speedup:.2f}x faster")
	else:
	print(f"✗ Custom implementation is {1/speedup:.2f}x slower")

	print("=" * 80)

	# Optional: Save detailed timing data
	print(f"\nDetailed timings (ms):")
	print(f"Reference: {ref_times}")
	print(f"Custom: {custom_times}")


	def test_moe_forward_benchmark_memory():
	"""Benchmark memory usage during forward pass."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True)

	# Initialize parameters
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	# Set reference layer parameters
	ref_layer.gate_up_proj = gate_up_proj
	ref_layer.gate_up_proj_bias = gate_up_proj_bias
	ref_layer.down_proj = down_proj
	ref_layer.down_proj_bias = down_proj_bias

	# Measure memory for reference implementation
	torch.cuda.empty_cache()
	torch.cuda.reset_peak_memory_stats()

	with torch.no_grad():
	output = ref_layer(hidden_states, router_idx, router_wt)

	ref_memory = torch.cuda.max_memory_allocated() / 1024**3 # GB

	# Measure memory for custom implementation
	torch.cuda.empty_cache()
	torch.cuda.reset_peak_memory_stats()

	with torch.no_grad():
	output = layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)

	custom_memory = torch.cuda.max_memory_allocated() / 1024**3 # GB

	print("\n" + "=" * 60)
	print("FORWARD MEMORY USAGE BENCHMARK")
	print("=" * 60)
	print(f"Reference implementation: {ref_memory:.3f} GB")
	print(f"Custom implementation: {custom_memory:.3f} GB")
	print(f"Memory ratio: {custom_memory/ref_memory:.3f}x")

	if custom_memory < ref_memory:
	print(f"✓ Custom uses {(1 - custom_memory/ref_memory)*100:.1f}% less memory")
	else:
	print(f"✗ Custom uses {(custom_memory/ref_memory - 1)*100:.1f}% more memory")
	print("=" * 60)


	def test_moe_forward_benchmark_throughput():
	"""Benchmark throughput (tokens/second) for forward pass."""

	# Test configuration
	num_experts = 128
	hidden_size = 1024
	expert_dim = 512
	batch_tokens = 4096
	topk = 2
	num_runs = 100

	torch.manual_seed(1337)
	torch.cuda.manual_seed(1337)

	# Generate test data
	hidden_states = torch.randn(
	batch_tokens,
	hidden_size,
	device="cuda",
	dtype=torch.float32,
	)
	router_idx = torch.randint(0, num_experts, (batch_tokens, topk), device="cuda")
	router_wt = torch.rand(batch_tokens, topk, device="cuda")
	router_wt = router_wt / router_wt.sum(dim=-1, keepdim=True)

	# Initialize parameters
	gate_up_proj = nn.Parameter(
	torch.randn(num_experts, hidden_size, 2 * expert_dim, device="cuda")
	)
	gate_up_proj_bias = nn.Parameter(
	torch.randn(num_experts, 2 * expert_dim, device="cuda")
	)
	down_proj = nn.Parameter(
	torch.randn(num_experts, expert_dim, hidden_size, device="cuda")
	)
	down_proj_bias = nn.Parameter(torch.randn(num_experts, hidden_size, device="cuda"))
	alpha = 1.702

	# Create Layers
	layer = MoE()
	ref_layer = OpenaiExperts(
	config=type(
	"Config",
	(object,),
	{
	"num_local_experts": num_experts,
	"intermediate_size": expert_dim,
	"hidden_size": hidden_size,
	},
	)
	)

	# Set reference layer parameters
	ref_layer.gate_up_proj = gate_up_proj
	ref_layer.gate_up_proj_bias = gate_up_proj_bias
	ref_layer.down_proj = down_proj
	ref_layer.down_proj_bias = down_proj_bias

	# Warmup
	print("Warming up for throughput test...")
	for _ in range(10):
	with torch.no_grad():
	ref_layer(hidden_states, router_idx, router_wt)
	layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)

	# Benchmark reference throughput
	torch.cuda.synchronize()
	start_time = time.perf_counter()

	for _ in range(num_runs):
	with torch.no_grad():
	ref_layer(hidden_states, router_idx, router_wt)

	torch.cuda.synchronize()
	ref_total_time = time.perf_counter() - start_time

	# Benchmark custom throughput
	torch.cuda.synchronize()
	start_time = time.perf_counter()

	for _ in range(num_runs):
	with torch.no_grad():
	layer(
	hidden_states,
	router_idx,
	router_wt,
	alpha,
	gate_up_proj,
	gate_up_proj_bias,
	down_proj,
	down_proj_bias,
	)

	torch.cuda.synchronize()
	custom_total_time = time.perf_counter() - start_time

	# Calculate throughput
	total_tokens_processed = batch_tokens * num_runs
	ref_throughput = total_tokens_processed / ref_total_time
	custom_throughput = total_tokens_processed / custom_total_time

	print("\n" + "=" * 70)
	print("FORWARD THROUGHPUT BENCHMARK")
	print("=" * 70)
	print(
	f"Configuration: {batch_tokens} tokens/batch × {num_runs} runs = {total_tokens_processed:,} tokens"
	)
	print()
	print(f"Reference Implementation:")
	print(f" - Total time: {ref_total_time:.3f} seconds")
	print(f" - Throughput: {ref_throughput:,.0f} tokens/second")
	print()
	print(f"Custom Implementation:")
	print(f" - Total time: {custom_total_time:.3f} seconds")
	print(f" - Throughput: {custom_throughput:,.0f} tokens/second")
	print()
	print(f"Throughput improvement: {custom_throughput/ref_throughput:.2f}x")

	if custom_throughput > ref_throughput:
	print(
	f"✓ Custom processes {((custom_throughput/ref_throughput - 1)*100):.1f}% more tokens/second"
	)
	else:
	print(
	f"✗ Custom processes {((1 - custom_throughput/ref_throughput)*100):.1f}% fewer tokens/second"
	)
	print("=" * 70)