zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-12-16 01:49:05 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Add quantize/dequantize for mxfp8 and nvfp4 (#2688)

Browse Source 
* Add quantize/dequantize slow path for mxfp8 and nvfp4

* fast cuda kernel for mx/nv quantization

* fallback for cuda < 12.8 (#2697)

* format (#2700)

* fix (#2701)

* metal kernels

* docs

* fix jit

* add default bits and group sizes

* improve quant docs

* fix output type of mxfp4 matmuls
This commit is contained in:
Awni Hannun
2025-10-28 16:23:12 -07:00
committed by GitHub
parent 460691a0e8
commit ec72b44417
25 changed files with 1400 additions and 588 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
mlx/backend/metal/CMakeLists.txt
View File

@@ -29,7 +29,7 @@ make_jit_source(
kernels/bf16_math.h
kernels/complex.h
kernels/defines.h)
make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h)
make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h kernels/fp8.h)
make_jit_source(binary_ops)
make_jit_source(ternary_ops)
make_jit_source(reduce_utils kernels/atomic.h kernels/reduction/ops.h)
@@ -81,7 +81,8 @@ if(MLX_METAL_JIT)
make_jit_source(quantized_utils)
make_jit_source(quantized kernels/quantized_utils.h)
make_jit_source(fp4_quantized kernels/quantized_utils.h)
make_jit_source(fp_quantized kernels/quantized_utils.h kernels/fp8.h
kernels/fp4.h)
make_jit_source(gemv_masked)
else()
target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/nojit_kernels.cpp)