pymllm.quantization.kernels.int8_activation_triton
==================================================

.. py:module:: pymllm.quantization.kernels.int8_activation_triton

.. autoapi-nested-parse::

   Per-token INT8 activation quantization using Triton.

   Ported from sglang int8_kernel.py (per_token_quant_int8).
   Original: sglang/srt/layers/quantization/int8_kernel.py:28-89



Functions
---------

.. autoapisummary::

   pymllm.quantization.kernels.int8_activation_triton.per_token_quant_int8


Module Contents
---------------

.. py:function:: per_token_quant_int8(x, scale_dtype = torch.float32)

   Per-token dynamic INT8 quantization.

   :param x: Input tensor, any shape with last dim = hidden_dim. Must be contiguous.
   :param scale_dtype: Dtype for scale output (default float32).

   :returns: INT8 quantized tensor, same shape as x.
             scales: Per-token scales, shape = x.shape[:-1] + (1,).
   :rtype: x_q