Skip to content

[NPU]: optimize GEGLU implementation with flatten 1D approach #1031

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
noemotiovon wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[NPU]: optimize GEGLU implementation with flatten 1D approach #1031

noemotiovon wants to merge 1 commit into from
+121 −195

Conversation

@noemotiovon
Copy link
Contributor

@noemotiovon noemotiovon commented Jan 20, 2026

  • Refactor Ascend GEGLU kernels to use flatten 1D grid-stride loop pattern instead of row-based tiling approach for better performance
  • Simplify block size calculation using compute_default_tiling_strategy
  • Align type conversion logic with GPU version for consistency
  • Update test tolerances for NPU bfloat16 (1e4) to handle precision differences

Hardware Type: Ascend 910B4

  • run make test to ensure correctness
  • run make checkstyle to ensure code style
  • run make test-convergence to ensure convergence

@noemotiovon
Copy link
Contributor Author

noemotiovon commented Jan 20, 2026

Benchmark:

**************************************
     BENCHMARKING SPEED for GEGLU
**************************************
[WARNING] Please DO NOT tune args ['num_warps']!
[WARNING] Please DO NOT tune args ['num_warps']!
********** Benchmark Data **********
[
  {
    "kernel_name": "geglu",
    "kernel_provider": "liger",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      41.29764175415039,
      77.97412109375,
      153.22802734375,
      305.4518127441406
    ],
    "y_values_20": [
      41.29764175415039,
      77.97412109375,
      153.22802734375,
      305.4518127441406
    ],
    "y_values_80": [
      41.29764175415039,
      77.97412109375,
      153.22802734375,
      305.4518127441406
    ],
    "timestamp": "2026-01-20 02:12:01",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "huggingface",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      43.27560043334961,
      82.03520202636719,
      159.79339599609375,
      318.13525390625
    ],
    "y_values_20": [
      43.27560043334961,
      82.03520202636719,
      159.79339599609375,
      318.13525390625
    ],
    "y_values_80": [
      43.27560043334961,
      82.03520202636719,
      159.79339599609375,
      318.13525390625
    ],
    "timestamp": "2026-01-20 02:12:10",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "liger",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      10.68362045288086,
      20.787700653076172,
      41.33747863769531,
      82.21530151367188
    ],
    "y_values_20": [
      10.68362045288086,
      20.787700653076172,
      41.33747863769531,
      82.21530151367188
    ],
    "y_values_80": [
      10.68362045288086,
      20.787700653076172,
      41.33747863769531,
      82.21530151367188
    ],
    "timestamp": "2026-01-20 02:12:15",
    "kernel_operation_mode": "forward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "huggingface",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      11.14247989654541,
      21.72364044189453,
      43.50083923339844,
      86.65560150146484
    ],
    "y_values_20": [
      11.14247989654541,
      21.72364044189453,
      43.50083923339844,
      86.65560150146484
    ],
    "y_values_80": [
      11.14247989654541,
      21.72364044189453,
      43.50083923339844,
      86.65560150146484
    ],
    "timestamp": "2026-01-20 02:12:20",
    "kernel_operation_mode": "forward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "liger",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      27.273399353027344,
      50.27893829345703,
      98.00166320800781,
      193.2967987060547
    ],
    "y_values_20": [
      27.273399353027344,
      50.27893829345703,
      98.00166320800781,
      193.2967987060547
    ],
    "y_values_80": [
      27.273399353027344,
      50.27893829345703,
      98.00166320800781,
      193.2967987060547
    ],
    "timestamp": "2026-01-20 02:12:27",
    "kernel_operation_mode": "backward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "huggingface",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      28.78070068359375,
      53.18796157836914,
      102.28961944580078,
      204.07730102539062
    ],
    "y_values_20": [
      28.78070068359375,
      53.18796157836914,
      102.28961944580078,
      204.07730102539062
    ],
    "y_values_80": [
      28.78070068359375,
      53.18796157836914,
      102.28961944580078,
      204.07730102539062
    ],
    "timestamp": "2026-01-20 02:12:35",
    "kernel_operation_mode": "backward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  }
]
**************************************
     BENCHMARKING MEMORY for GEGLU
**************************************
********** Benchmark Data **********
[
  {
    "kernel_name": "geglu",
    "kernel_provider": "liger",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      1910.0087890625,
      3092.0078125,
      5668.0078125,
      10820.0078125
    ],
    "y_values_20": [
      1910.0087890625,
      3092.0078125,
      5668.0078125,
      10820.0078125
    ],
    "y_values_80": [
      1910.0087890625,
      3092.0078125,
      5668.0078125,
      10820.0078125
    ],
    "timestamp": "2026-01-20 02:12:40",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "huggingface",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      2082.00927734375,
      3436.00830078125,
      6356.00830078125,
      12196.0078125
    ],
    "y_values_20": [
      2082.00927734375,
      3436.00830078125,
      6356.00830078125,
      12196.0078125
    ],
    "y_values_80": [
      2082.00927734375,
      3436.00830078125,
      6356.00830078125,
      12196.0078125
    ],
    "timestamp": "2026-01-20 02:12:49",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "liger",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      922.0048828125,
      1566.0048828125,
      2854.0048828125,
      5430.0048828125
    ],
    "y_values_20": [
      922.0048828125,
      1566.0048828125,
      2854.0048828125,
      5430.0048828125
    ],
    "y_values_80": [
      922.0048828125,
      1566.0048828125,
      2854.0048828125,
      5430.0048828125
    ],
    "timestamp": "2026-01-20 02:12:56",
    "kernel_operation_mode": "forward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "huggingface",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      1094.00537109375,
      1910.00537109375,
      3542.00537109375,
      6806.00537109375
    ],
    "y_values_20": [
      1094.00537109375,
      1910.00537109375,
      3542.00537109375,
      6806.00537109375
    ],
    "y_values_80": [
      1094.00537109375,
      1910.00537109375,
      3542.00537109375,
      6806.00537109375
    ],
    "timestamp": "2026-01-20 02:13:00",
    "kernel_operation_mode": "forward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "liger",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      1910.0087890625,
      3092.0078125,
      5668.0078125,
      10820.0078125
    ],
    "y_values_20": [
      1910.0087890625,
      3092.0078125,
      5668.0078125,
      10820.0078125
    ],
    "y_values_80": [
      1910.0087890625,
      3092.0078125,
      5668.0078125,
      10820.0078125
    ],
    "timestamp": "2026-01-20 02:13:04",
    "kernel_operation_mode": "backward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  },
  {
    "kernel_name": "geglu",
    "kernel_provider": "huggingface",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "T",
    "x_label": "sequence length",
    "x_values": [
      1024,
      2048,
      4096,
      8192
    ],
    "y_values_50": [
      2082.00927734375,
      3436.00830078125,
      6356.00830078125,
      12196.0078125
    ],
    "y_values_20": [
      2082.00927734375,
      3436.00830078125,
      6356.00830078125,
      12196.0078125
    ],
    "y_values_80": [
      2082.00927734375,
      3436.00830078125,
      6356.00830078125,
      12196.0078125
    ],
    "timestamp": "2026-01-20 02:13:11",
    "kernel_operation_mode": "backward",
    "extra_benchmark_config_str": "{\"bsz\": 8, \"hidden_size\": 4096, \"intermediate_size\": 11008, \"hidden_act\": \"gelu_pytorch_tanh\", \"dtype\": \"torch.bfloat16\"}",
    "liger_version": "0.6.4"
  }
]

@noemotiovon
[NPU]: optimize GEGLU implementation with flatten 1D approach
b524da8 
- Refactor Ascend GEGLU kernels to use flatten 1D grid-stride loop pattern
  instead of row-based tiling approach for better performance
- Simplify block size calculation using compute_default_tiling_strategy
- Align type conversion logic with GPU version for consistency
- Update test tolerances for NPU bfloat16 (1e4) to handle precision differences
Tcc0403
Tcc0403 reviewed Jan 20, 2026
View reviewed changes
test/transformers/test_geglu.py
Comment on lines +39 to +40
# TODO: we should find a better way to tune this. 1e4 is too large apparently
1e-2 if device != "npu" else 1e4,
Copy link
Collaborator

@Tcc0403 Tcc0403 Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you know what tensor couldn't pass with this tolerance? gradients or inputs?

Copy link
Contributor Author

@noemotiovon noemotiovon Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the question. I double-checked which tensors require the large tolerance.

On NPU with bfloat16:

  • Forward outputs (y1 vs y2) differ at around O(1e2).
  • Weight gradients (gate_proj / up_proj / down_proj) are also at O(1e2).
  • The largest discrepancy is in the input gradients: x1.grad vs x2.grad can reach O(1e4).

So the forward and weight gradients are already numerically different at ~1e2, and the input gradients further amplify this difference.

Copy link
Contributor Author

@noemotiovon noemotiovon Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more. 

================================================================================
SUMMARY - Minimum atol needed for each tensor (rtol=1e-2):
================================================================================
output                        : min_atol=1e2   , max_abs_diff=2.048000e+03
gate_proj.weight.grad         : min_atol=1e3   , max_abs_diff=2.048000e+03
up_proj.weight.grad           : min_atol=1e2   , max_abs_diff=2.048000e+03
down_proj.weight.grad         : min_atol=1e2   , max_abs_diff=2.048000e+03
input.grad                    : min_atol=1e4   , max_abs_diff=4.096000e+03

Copy link
Contributor Author

@noemotiovon noemotiovon Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also worth noting: the tolerances used here are consistent with the previous NPU GEGLU kernel implementation, so this change does not introduce new numerical error compared to the existing behavior on NPU.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Tcc0403 Tcc0403 Tcc0403 left review comments

At least 1 approving review is required to merge this pull request.

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@noemotiovon @Tcc0403