A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 09:30 UTC
[BugFix][Kernel] Fix Illegal memory access in causal_conv1d in H100 (#9838)
Signed-off-by: mzusman <mor.zusmann@gmail.com>
Mor Zusman committed
Oct 31, 2024 at 22:06 UTC
9fb12f7848d427b6c1c29052271030a5e96bd74a
3 files changed
+40
-7
csrc/mamba/causal_conv1d/causal_conv1d.cu
+32
-2
@@ -418,6 +418,31 @@ void causal_conv1d_fwd_kernel(ConvParamsBase params) {
418
typename Ktraits::BlockStoreT(smem_store).Store(out, out_vals_store, seqlen - chunk * kChunkSize);
419
}
420
out += kChunkSize;
421
+
422
+ int final_state_position = ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize);
423
+ // in case the final state is separated between the last "smem_exchange" and
424
+ // and the one before it (chunk = n_chunks - 1 and chunk = n_chunks - 2),
425
+ // (which occurs when `final_state_position` is a non-positivie index)
426
+ // we load the correct data from smem_exchange from both chunks, the last chunk iteration and the one before it
427
+ if (final_state_position < 0 && seqlen > kWidth){
428
+ input_t vals_load[kNElts] = {0};
429
+ if ((chunk == n_chunks - 2) && (tidx == kNThreads - 1)){
430
+ // chunk = n_chunks - 2, a segment of the final state sits in the last index
431
+ reinterpret_cast<vec_t *>(vals_load)[0] = smem_exchange[kNThreads - 1];
432
+ #pragma unroll
433
+ for (int w = 0; w < -final_state_position; ++w){
434
+ conv_states[w] = vals_load[kNElts + final_state_position + w];
435
+ }
436
+ }
437
+ if ((chunk == n_chunks - 1) && tidx == 0){
438
+ // chunk = n_chunks - 1, the second segment of the final state first positions
439
+ reinterpret_cast<vec_t *>(vals_load)[0] = smem_exchange[0];
440
+ for (int w = -final_state_position; w < kWidth - 1; ++w){
441
+ conv_states[w] = vals_load[w + final_state_position];
442
+ }
443
+ return;
444
+ }
445
+ }
446
}
447
// Final state is stored in the smem_exchange last token slot,
448
// in case seqlen < kWidth, we would need to take the final state from the
@@ -446,9 +471,14 @@ void causal_conv1d_fwd_kernel(ConvParamsBase params) {
471
}
472
else {
473
// in case the final state is in between the threads data
449
- reinterpret_cast<vec_t *>(x_vals_load)[1] = smem_exchange[last_thread + 1];
450
- reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[last_thread];
474
const int offset = ((seqlen - (kWidth - 1)) % (kNElts));
475
+ if ((offset + kWidth - 2) >= kNElts && (last_thread + 1 < kNThreads)){
476
+ // In case last_thread == kNThreads - 1, accessing last_thread + 1 will result in a
477
+ // illegal access error on H100.
478
+ // Therefore, we access last_thread + 1, only if the final state data sits there
479
+ reinterpret_cast<vec_t *>(x_vals_load)[1] = smem_exchange[last_thread + 1];
480
+ }
481
+ reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[last_thread];
482
#pragma unroll
483
for (int w = 0; w < kWidth - 1; ++w){
484
conv_states[w] = x_vals_load[offset + w ];
tests/kernels/test_causal_conv1d.py
+5
-2
@@ -151,7 +151,7 @@ def causal_conv1d_opcheck_fn(x: torch.Tensor,
151
@pytest.mark.parametrize("has_bias", [True])
152
@pytest.mark.parametrize("width", [4])
153
@pytest.mark.parametrize(
154
- 'seqlen', [1, 8, 16, 32, 64, 128, 256, 512, 784, 1024, 2048, 4096])
154
+ 'seqlen', [1, 8, 16, 32, 64, 128, 256, 512, 784, 1024, 1025, 2048, 4096])
155
@pytest.mark.parametrize('dim', [64])
156
@pytest.mark.parametrize('batch', [1])
157
def test_causal_conv1d(batch, dim, seqlen, width, has_bias, silu_activation,
@@ -420,7 +420,10 @@ def test_causal_conv1d_varlen(with_padding, dim, seqlen, width, has_bias,
420
421
unpadded_out = out[:, :out_ref_tensor.shape[-1]]
422
assert torch.allclose(unpadded_out, out_ref_tensor, rtol=rtol, atol=atol)
423
- assert torch.allclose(final_states, final_states_ref, rtol=rtol, atol=atol)
423
+ assert torch.allclose(final_states[state_indices],
424
+ final_states_ref[state_indices],
425
+ rtol=rtol,
426
+ atol=atol)
427
428
causal_conv1d_opcheck_fn(x.squeeze(0), weight, bias, cumsum.cuda(),
429
padded_state_indices, has_initial_states,
tests/kernels/test_mamba_ssm.py
+3
-3
@@ -555,7 +555,7 @@ def test_selective_state_update_with_batch_indices(with_padding, dim, dstate,
555
device = "cuda"
556
rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2)
557
if itype == torch.bfloat16:
558
- rtol, atol = 7e-2, 7e-2
558
+ rtol, atol = 1e-1, 1e-1
559
if torch.version.hip:
560
atol *= 2
561
# set seed
@@ -610,8 +610,8 @@ def test_selective_state_update_with_batch_indices(with_padding, dim, dstate,
610
dt_bias=dt_bias,
611
dt_softplus=True)
612
613
- print("Output diff max", (out - out_ref[0]).max())
614
- print("Output diff mean", (out - out_ref[0]).mean())
613
+ print("Output diff max", (out[:batch_size] - out_ref).max())
614
+ print("Output diff mean", (out[:batch_size] - out_ref).mean())
615
print("Output state diff max", (state[state_indices, :] - state_ref).max())
616
print("Output state diff mean",
617
(state[state_indices, :] - state_ref).mean())