fix-grid-limits

#2
by 3outeille HF Staff - opened
build/torch27-cxx11-cu118-x86_64-linux/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,
build/torch27-cxx11-cu126-x86_64-linux/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,
build/torch27-cxx11-cu128-x86_64-linux/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,
build/torch28-cxx11-cu126-x86_64-linux/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,
build/torch28-cxx11-cu128-x86_64-linux/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,
build/torch28-cxx11-cu129-x86_64-linux/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,
tests/ops/binned_copy.py ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import pytest
3
+
4
+ from megablocks.ops.binned_gather import BinnedGatherOp
5
+
6
+ binned_gather_triton = BinnedGatherOp.apply
7
+
8
+ def set_seeds(seed=0):
9
+ torch.manual_seed(seed)
10
+ if torch.cuda.is_available():
11
+ torch.cuda.manual_seed_all(seed)
12
+
13
+ # Stress test expert_capacity, especially near and at the upper limit (e.g., 65535 for int16 indexing)
14
+ def make_stress_expert_capacity_tests():
15
+ tests = []
16
+ # Small cases for sanity
17
+ for seq_len, hidden_size, num_experts, top_k in [
18
+ (4, 2, 2, 1),
19
+ (4, 2, 2, 2),
20
+ (4, 2, 2, 4),
21
+ ]:
22
+ for expert_capacity in [1, 2, 4]:
23
+ tests.append((seq_len, hidden_size, num_experts, top_k, expert_capacity))
24
+ # Medium cases
25
+ for seq_len, hidden_size, num_experts, top_k in [
26
+ (1024, 1536, 4, 1),
27
+ (1024, 1536, 4, 2),
28
+ (1024, 1536, 4, 4),
29
+ (1024, 1536, 64, 1),
30
+ (1024, 1536, 64, 2),
31
+ (1024, 1536, 64, 4),
32
+ (1024, 1536, 128, 1),
33
+ (1024, 1536, 128, 2),
34
+ (1024, 1536, 128, 4),
35
+ ]:
36
+ for expert_capacity in [1, 2, 4, 128, 1024]:
37
+ tests.append((seq_len, hidden_size, num_experts, top_k, expert_capacity))
38
+
39
+ # Large cases, stress expert_capacity near 65536 (CUDA second dim grid limit)
40
+ for seq_len, hidden_size, num_experts, top_k in [
41
+ (4096, 768, 32, 4),
42
+ ]:
43
+ for expert_capacity in [65535, 70000, 90000]:
44
+ tests.append((seq_len, hidden_size, num_experts, top_k, expert_capacity))
45
+
46
+ return tuple(tests)
47
+
48
+ BINNED_GATHER_TESTS = make_stress_expert_capacity_tests()
49
+
50
+ @pytest.mark.parametrize(('seq_len', 'hidden_size', 'num_experts', 'top_k', 'expert_capacity'), BINNED_GATHER_TESTS)
51
+ def test_binned_gather(seq_len: int, hidden_size: int, num_experts: int, top_k: int, expert_capacity: int):
52
+ # NOTE: Capacity factor == 1.
53
+ set_seeds(42)
54
+ # Create the data and indices with gradient tracking
55
+ x = torch.arange(seq_len * hidden_size, device='cuda', dtype=torch.half).view(seq_len, hidden_size)
56
+ x.requires_grad_(True)
57
+
58
+ # Randomly assign tokens to experts.
59
+ top_expert = torch.randint(0, num_experts, (seq_len * top_k,), device='cuda', dtype=torch.int)
60
+ _, indices = torch.sort(top_expert)
61
+ bins = torch.cumsum(torch.bincount(top_expert, minlength=num_experts), 0).to(torch.int32)
62
+ # Example: counts is [12, 2, 3], the bins tensor will be [12, 14, 20]. This tells the gather function:
63
+ # Expert 0's assignments are in indices[0:12].
64
+ # Expert 1's assignments are in indices[12:14].
65
+ # Expert 2's assignments are in indices[14:20]. (we have num_tokens * 3)
66
+
67
+ def binned_gather_pytorch(
68
+ x: torch.Tensor,
69
+ indices: torch.Tensor,
70
+ bins: torch.Tensor,
71
+ expert_capacity: int,
72
+ top_k: int,
73
+ ):
74
+ start = 0
75
+ out = torch.zeros((num_experts, expert_capacity, hidden_size), dtype=x.dtype, device=x.device)
76
+ for i in range(num_experts):
77
+ end = bins[i]
78
+ num_tokens = min(expert_capacity, end - start)
79
+ if num_tokens > 0:
80
+ # indices[start:end] are the indices for this expert
81
+ # For each slot j, get the input index and copy the row
82
+ idx = indices[start : start + num_tokens] // top_k
83
+ out[i, :num_tokens, :] = x[idx, :]
84
+ start = end
85
+ return out
86
+
87
+ out = binned_gather_triton(x, indices, bins, expert_capacity, top_k)
88
+ expected_out = binned_gather_pytorch(x, indices, bins, expert_capacity, top_k)
89
+ assert torch.all(torch.eq(out, expected_out))
90
+
91
+ # Test backward pass
92
+ grad_output = torch.arange(out.numel(), device=out.device, dtype=out.dtype).view_as(out)
93
+ out.backward(grad_output)
94
+
95
+ # Verify gradients were computed
96
+ assert x.grad is not None, "Gradients should be computed for input x"
97
+ assert x.grad.shape == x.shape, f"Gradient shape {x.grad.shape} should match input shape {x.shape}"
98
+
99
+ # Reference implementation for backward pass (binned_scatter)
100
+ def binned_scatter_pytorch(
101
+ x: torch.Tensor,
102
+ indices: torch.Tensor,
103
+ weights: torch.Tensor,
104
+ bins: torch.Tensor,
105
+ top_k: int,
106
+ ):
107
+ # x: (ne, ec, hs)
108
+ # indices: (sl * top_k,)
109
+ # weights: (sl * top_k,)
110
+ # bins: (ne,)
111
+ # Output: (sl, hs)
112
+ out = torch.zeros((seq_len, hidden_size), device=x.device, dtype=x.dtype)
113
+ start = 0
114
+ for i in range(num_experts):
115
+ end = bins[i]
116
+ num_tokens = min(expert_capacity, end - start)
117
+ for j in range(num_tokens):
118
+ index = indices[start + j]
119
+ scale = weights[index] if weights is not None else 1.0
120
+ token_pos = index // top_k
121
+
122
+ out[token_pos, :] += scale * x[i, j, :]
123
+ start = end
124
+ return out
125
+
126
+ expected_grad = binned_scatter_pytorch(grad_output, indices, None, bins, top_k)
127
+ print(f"x.grad: {x.grad}")
128
+ print(f"expected_grad: {expected_grad}")
129
+
130
+ # Use torch.allclose instead of exact equality for floating point comparison
131
+ if torch.allclose(x.grad, expected_grad, rtol=1e-3, atol=1e-3):
132
+ print("✅ Success: Gradients match!")
133
+ else:
134
+ print("❌ Gradients don't match")
135
+ # Let's see if it's just a reordering issue
136
+ print("Checking if values match when sorted...")
137
+ grad_sorted = torch.sort(x.grad.flatten())[0]
138
+ expected_sorted = torch.sort(expected_grad.flatten())[0]
139
+ if torch.allclose(grad_sorted, expected_sorted, rtol=1e-3, atol=1e-3):
140
+ print("✅ Same values, different order - routing issue!")
141
+ else:
142
+ print("❌ Different values entirely")
143
+
144
+ print(f"\nTriton Output Shape: {x.grad.shape}")
145
+ print(f"PyTorch Output Shape: {expected_grad.shape}")
torch-ext/megablocks/backend/kernels.py CHANGED
@@ -352,8 +352,8 @@ def _binned_copy(
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
- expert_idx = tl.program_id(0)
356
- entry_idx = tl.program_id(1)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
@@ -416,7 +416,7 @@ def binned_gather(x, indices, weights, bins, expert_capacity, top_k):
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
- _binned_copy[(num_experts, expert_capacity)](
420
  x,
421
  out,
422
  num_experts,
@@ -445,7 +445,7 @@ def binned_scatter(x, indices, weights, bins, top_k):
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
- _binned_copy[(num_experts, expert_capacity)](
449
  out,
450
  x,
451
  num_experts,
@@ -492,8 +492,8 @@ def _binned_copy_wgrad(
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
- expert_idx = tl.program_id(0)
496
- entry_idx = tl.program_id(1)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
@@ -543,7 +543,7 @@ def binned_scatter_wgrad(x, grad, indices, bins, top_k):
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
- _binned_copy_wgrad[(num_experts, expert_capacity)](
547
  x,
548
  grad,
549
  out,
 
352
  SCALE: tl.constexpr,
353
  ):
354
  # Load our indices into the output.
355
+ expert_idx = tl.program_id(1)
356
+ entry_idx = tl.program_id(0)
357
 
358
  # Calculate our offset into the output.
359
  index_b = expert_idx * expert_capacity + entry_idx
 
416
  num_experts = bins.shape[0]
417
  out = torch.zeros((num_experts, expert_capacity, x.shape[1]), dtype=x.dtype, device=x.device)
418
 
419
+ _binned_copy[(expert_capacity, num_experts)](
420
  x,
421
  out,
422
  num_experts,
 
445
  num_experts, expert_capacity, hidden_size = x.shape
446
  tokens = indices.shape[0] // top_k
447
  out = torch.zeros((tokens, top_k, hidden_size), dtype=x.dtype, device=x.device)
448
+ _binned_copy[(expert_capacity, num_experts)](
449
  out,
450
  x,
451
  num_experts,
 
492
  BLOCK_X: tl.constexpr,
493
  ):
494
  # Load our indices into the output.
495
+ expert_idx = tl.program_id(1)
496
+ entry_idx = tl.program_id(0)
497
 
498
  # Calculate our offset into the output.
499
  index_x = expert_idx * expert_capacity + entry_idx
 
543
  num_experts, expert_capacity, hidden_size = x.shape
544
  tokens = indices.shape[0] // top_k
545
  out = torch.zeros((tokens * top_k), dtype=x.dtype, device=x.device)
546
+ _binned_copy_wgrad[(expert_capacity, num_experts)](
547
  x,
548
  grad,
549
  out,