mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-05-04 21:36:14 +02:00
be95871adc
Some checks failed
Python Linting / Run Ruff (push) Has been cancelled
Python Linting / Run Pylint (push) Has been cancelled
Full Comfy CI Workflow Runs / test-stable (12.1, , linux, 3.10, [self-hosted Linux], stable) (push) Has been cancelled
Full Comfy CI Workflow Runs / test-stable (12.1, , linux, 3.11, [self-hosted Linux], stable) (push) Has been cancelled
Full Comfy CI Workflow Runs / test-stable (12.1, , linux, 3.12, [self-hosted Linux], stable) (push) Has been cancelled
Full Comfy CI Workflow Runs / test-unix-nightly (12.1, , linux, 3.11, [self-hosted Linux], nightly) (push) Has been cancelled
Execution Tests / test (macos-latest) (push) Has been cancelled
Execution Tests / test (ubuntu-latest) (push) Has been cancelled
Execution Tests / test (windows-latest) (push) Has been cancelled
Test server launches without errors / test (push) Has been cancelled
Unit Tests / test (macos-latest) (push) Has been cancelled
Unit Tests / test (ubuntu-latest) (push) Has been cancelled
Unit Tests / test (windows-2022) (push) Has been cancelled
Build package / Build Test (3.10) (push) Has been cancelled
Build package / Build Test (3.11) (push) Has been cancelled
Build package / Build Test (3.12) (push) Has been cancelled
Build package / Build Test (3.13) (push) Has been cancelled
Build package / Build Test (3.14) (push) Has been cancelled
* initial gemma4 support * parity with reference implementation outputs can 100% match transformers with same sdpa flags, checkpoint this and then optimize * Cleanup, video fixes * cleanup, enable fused rms norm by default * update comment * Cleanup * Update sd.py * Various fixes * Add fp8 scaled embedding support * small fixes * Translate think tokens * Fix image encoder attention mask type So it works with basic attention * Handle thinking tokens different only for Gemma4 * Code cleanup * Update nodes_textgen.py * Use embed scale class instead of buffer Slight difference to HF, but technically more accurate and simpler code * Default to fused rms_norm * Update gemma4.py
12 lines
504 B
Python
12 lines
504 B
Python
import torch
|
|
import comfy.model_management
|
|
|
|
RMSNorm = torch.nn.RMSNorm
|
|
|
|
# Note: torch's fused F.rms_norm is faster but produces slightly different output than manual implementations (rsqrt/reduction rounding).
|
|
def rms_norm(x, weight=None, eps=1e-6):
|
|
if weight is None:
|
|
return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps)
|
|
else:
|
|
return torch.nn.functional.rms_norm(x, weight.shape, weight=comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)
|