diff --git a/MLX_CORE_VERSION b/MLX_CORE_VERSION index 912750052..5affb131c 100644 --- a/MLX_CORE_VERSION +++ b/MLX_CORE_VERSION @@ -1 +1 @@ -v0.30.6 +v0.31.1 diff --git a/MLX_VERSION b/MLX_VERSION index b043aa648..60f634328 100644 --- a/MLX_VERSION +++ b/MLX_VERSION @@ -1 +1 @@ -v0.5.0 +v0.6.0 diff --git a/x/imagegen/mlx/mlx.c b/x/imagegen/mlx/mlx.c index b0ccbacdf..8d7ec0e0a 100644 --- a/x/imagegen/mlx/mlx.c +++ b/x/imagegen/mlx/mlx.c @@ -165,8 +165,8 @@ int (*mlx_distributed_sum_scatter_ptr)(mlx_array* res, const mlx_array x, const int (*mlx_distributed_group_rank_ptr)(mlx_distributed_group group) = NULL; int (*mlx_distributed_group_size_ptr)(mlx_distributed_group group) = NULL; mlx_distributed_group (*mlx_distributed_group_split_ptr)(mlx_distributed_group group, int color, int key) = NULL; -bool (*mlx_distributed_is_available_ptr)(void) = NULL; -mlx_distributed_group (*mlx_distributed_init_ptr)(bool strict) = NULL; +bool (*mlx_distributed_is_available_ptr)(const char* bk) = NULL; +mlx_distributed_group (*mlx_distributed_init_ptr)(bool strict, const char* bk) = NULL; void (*mlx_set_error_handler_ptr)(mlx_error_handler_func handler, void* data, void (*dtor)(void*)) = NULL; void (*_mlx_error_ptr)(const char* file, const int line, const char* fmt, ...) = NULL; int (*mlx_export_function_ptr)(const char* file, const mlx_closure fun, const mlx_vector_array args, bool shapeless) = NULL; @@ -319,10 +319,12 @@ int (*mlx_astype_ptr)(mlx_array* res, const mlx_array a, mlx_dtype dtype, const int (*mlx_atleast_1d_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_atleast_2d_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_atleast_3d_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; +int (*mlx_bartlett_ptr)(mlx_array* res, int M, const mlx_stream s) = NULL; int (*mlx_bitwise_and_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; int (*mlx_bitwise_invert_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_bitwise_or_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; int (*mlx_bitwise_xor_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; +int (*mlx_blackman_ptr)(mlx_array* res, int M, const mlx_stream s) = NULL; int (*mlx_block_masked_mm_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, int block_size, const mlx_array mask_out , const mlx_array mask_lhs , const mlx_array mask_rhs , const mlx_stream s) = NULL; int (*mlx_broadcast_arrays_ptr)(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_stream s) = NULL; int (*mlx_broadcast_to_ptr)(mlx_array* res, const mlx_array a, const int* shape, size_t shape_num, const mlx_stream s) = NULL; @@ -348,7 +350,7 @@ int (*mlx_cumprod_ptr)(mlx_array* res, const mlx_array a, int axis, bool reverse int (*mlx_cumsum_ptr)(mlx_array* res, const mlx_array a, int axis, bool reverse, bool inclusive, const mlx_stream s) = NULL; int (*mlx_degrees_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_depends_ptr)(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_vector_array dependencies) = NULL; -int (*mlx_dequantize_ptr)(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, mlx_optional_dtype dtype, const mlx_stream s) = NULL; +int (*mlx_dequantize_ptr)(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , mlx_optional_dtype dtype, const mlx_stream s) = NULL; int (*mlx_diag_ptr)(mlx_array* res, const mlx_array a, int k, const mlx_stream s) = NULL; int (*mlx_diagonal_ptr)(mlx_array* res, const mlx_array a, int offset, int axis1, int axis2, const mlx_stream s) = NULL; int (*mlx_divide_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; @@ -375,6 +377,8 @@ int (*mlx_gather_qmm_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, int (*mlx_greater_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; int (*mlx_greater_equal_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; int (*mlx_hadamard_transform_ptr)(mlx_array* res, const mlx_array a, mlx_optional_float scale, const mlx_stream s) = NULL; +int (*mlx_hamming_ptr)(mlx_array* res, int M, const mlx_stream s) = NULL; +int (*mlx_hanning_ptr)(mlx_array* res, int M, const mlx_stream s) = NULL; int (*mlx_identity_ptr)(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s) = NULL; int (*mlx_imag_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_inner_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; @@ -434,8 +438,8 @@ int (*mlx_prod_axes_ptr)(mlx_array* res, const mlx_array a, const int* axes, siz int (*mlx_prod_axis_ptr)(mlx_array* res, const mlx_array a, int axis, bool keepdims, const mlx_stream s) = NULL; int (*mlx_prod_ptr)(mlx_array* res, const mlx_array a, bool keepdims, const mlx_stream s) = NULL; int (*mlx_put_along_axis_ptr)(mlx_array* res, const mlx_array a, const mlx_array indices, const mlx_array values, int axis, const mlx_stream s) = NULL; -int (*mlx_qqmm_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s) = NULL; -int (*mlx_quantize_ptr)(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s) = NULL; +int (*mlx_qqmm_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale_x , const mlx_array global_scale_w , const mlx_stream s) = NULL; +int (*mlx_quantize_ptr)(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , const mlx_stream s) = NULL; int (*mlx_quantized_matmul_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array scales, const mlx_array biases , bool transpose, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s) = NULL; int (*mlx_radians_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_real_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; @@ -2101,6 +2105,11 @@ int mlx_load_functions(void* handle) { fprintf(stderr, "MLX: Failed to load symbol: mlx_atleast_3d\n"); return -1; } + mlx_bartlett_ptr = GET_SYM(handle, "mlx_bartlett"); + if (mlx_bartlett_ptr == NULL) { + fprintf(stderr, "MLX: Failed to load symbol: mlx_bartlett\n"); + return -1; + } mlx_bitwise_and_ptr = GET_SYM(handle, "mlx_bitwise_and"); if (mlx_bitwise_and_ptr == NULL) { fprintf(stderr, "MLX: Failed to load symbol: mlx_bitwise_and\n"); @@ -2121,6 +2130,11 @@ int mlx_load_functions(void* handle) { fprintf(stderr, "MLX: Failed to load symbol: mlx_bitwise_xor\n"); return -1; } + mlx_blackman_ptr = GET_SYM(handle, "mlx_blackman"); + if (mlx_blackman_ptr == NULL) { + fprintf(stderr, "MLX: Failed to load symbol: mlx_blackman\n"); + return -1; + } mlx_block_masked_mm_ptr = GET_SYM(handle, "mlx_block_masked_mm"); if (mlx_block_masked_mm_ptr == NULL) { fprintf(stderr, "MLX: Failed to load symbol: mlx_block_masked_mm\n"); @@ -2381,6 +2395,16 @@ int mlx_load_functions(void* handle) { fprintf(stderr, "MLX: Failed to load symbol: mlx_hadamard_transform\n"); return -1; } + mlx_hamming_ptr = GET_SYM(handle, "mlx_hamming"); + if (mlx_hamming_ptr == NULL) { + fprintf(stderr, "MLX: Failed to load symbol: mlx_hamming\n"); + return -1; + } + mlx_hanning_ptr = GET_SYM(handle, "mlx_hanning"); + if (mlx_hanning_ptr == NULL) { + fprintf(stderr, "MLX: Failed to load symbol: mlx_hanning\n"); + return -1; + } mlx_identity_ptr = GET_SYM(handle, "mlx_identity"); if (mlx_identity_ptr == NULL) { fprintf(stderr, "MLX: Failed to load symbol: mlx_identity\n"); @@ -4132,12 +4156,12 @@ mlx_distributed_group mlx_distributed_group_split(mlx_distributed_group group, i return mlx_distributed_group_split_ptr(group, color, key); } -bool mlx_distributed_is_available(void) { - return mlx_distributed_is_available_ptr(); +bool mlx_distributed_is_available(const char* bk) { + return mlx_distributed_is_available_ptr(bk); } -mlx_distributed_group mlx_distributed_init(bool strict) { - return mlx_distributed_init_ptr(strict); +mlx_distributed_group mlx_distributed_init(bool strict, const char* bk) { + return mlx_distributed_init_ptr(strict, bk); } void mlx_set_error_handler(mlx_error_handler_func handler, void* data, void (*dtor)(void*)) { @@ -4748,6 +4772,10 @@ int mlx_atleast_3d(mlx_array* res, const mlx_array a, const mlx_stream s) { return mlx_atleast_3d_ptr(res, a, s); } +int mlx_bartlett(mlx_array* res, int M, const mlx_stream s) { + return mlx_bartlett_ptr(res, M, s); +} + int mlx_bitwise_and(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s) { return mlx_bitwise_and_ptr(res, a, b, s); } @@ -4764,6 +4792,10 @@ int mlx_bitwise_xor(mlx_array* res, const mlx_array a, const mlx_array b, const return mlx_bitwise_xor_ptr(res, a, b, s); } +int mlx_blackman(mlx_array* res, int M, const mlx_stream s) { + return mlx_blackman_ptr(res, M, s); +} + int mlx_block_masked_mm(mlx_array* res, const mlx_array a, const mlx_array b, int block_size, const mlx_array mask_out , const mlx_array mask_lhs , const mlx_array mask_rhs , const mlx_stream s) { return mlx_block_masked_mm_ptr(res, a, b, block_size, mask_out, mask_lhs, mask_rhs, s); } @@ -4864,8 +4896,8 @@ int mlx_depends(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_ return mlx_depends_ptr(res, inputs, dependencies); } -int mlx_dequantize(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, mlx_optional_dtype dtype, const mlx_stream s) { - return mlx_dequantize_ptr(res, w, scales, biases, group_size, bits, mode, dtype, s); +int mlx_dequantize(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , mlx_optional_dtype dtype, const mlx_stream s) { + return mlx_dequantize_ptr(res, w, scales, biases, group_size, bits, mode, global_scale, dtype, s); } int mlx_diag(mlx_array* res, const mlx_array a, int k, const mlx_stream s) { @@ -4972,6 +5004,14 @@ int mlx_hadamard_transform(mlx_array* res, const mlx_array a, mlx_optional_float return mlx_hadamard_transform_ptr(res, a, scale, s); } +int mlx_hamming(mlx_array* res, int M, const mlx_stream s) { + return mlx_hamming_ptr(res, M, s); +} + +int mlx_hanning(mlx_array* res, int M, const mlx_stream s) { + return mlx_hanning_ptr(res, M, s); +} + int mlx_identity(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s) { return mlx_identity_ptr(res, n, dtype, s); } @@ -5208,12 +5248,12 @@ int mlx_put_along_axis(mlx_array* res, const mlx_array a, const mlx_array indice return mlx_put_along_axis_ptr(res, a, indices, values, axis, s); } -int mlx_qqmm(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s) { - return mlx_qqmm_ptr(res, x, w, w_scales, group_size, bits, mode, s); +int mlx_qqmm(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale_x , const mlx_array global_scale_w , const mlx_stream s) { + return mlx_qqmm_ptr(res, x, w, w_scales, group_size, bits, mode, global_scale_x, global_scale_w, s); } -int mlx_quantize(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s) { - return mlx_quantize_ptr(res, w, group_size, bits, mode, s); +int mlx_quantize(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , const mlx_stream s) { + return mlx_quantize_ptr(res, w, group_size, bits, mode, global_scale, s); } int mlx_quantized_matmul(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array scales, const mlx_array biases , bool transpose, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s) { diff --git a/x/imagegen/mlx/mlx.go b/x/imagegen/mlx/mlx.go index dc925570b..545f65b14 100644 --- a/x/imagegen/mlx/mlx.go +++ b/x/imagegen/mlx/mlx.go @@ -2124,8 +2124,9 @@ func Quantize(w *Array, groupSize, bits int, mode string) (weights, scales, bias defer C.free(unsafe.Pointer(cMode)) optGroupSize := C.mlx_optional_int{value: C.int(groupSize), has_value: true} optBits := C.mlx_optional_int{value: C.int(bits), has_value: true} + var globalScale C.mlx_array res := C.mlx_vector_array_new() - C.mlx_quantize(&res, w.c, optGroupSize, optBits, cMode, C.default_stream()) + C.mlx_quantize(&res, w.c, optGroupSize, optBits, cMode, globalScale, C.default_stream()) // Result is a vector of arrays: [weights, scales, biases?] // mxfp8 mode returns only 2 elements (no biases) @@ -2154,6 +2155,7 @@ func Dequantize(w, scales, biases *Array, groupSize, bits int, mode string) *Arr optGroupSize := C.mlx_optional_int{value: C.int(groupSize), has_value: true} optBits := C.mlx_optional_int{value: C.int(bits), has_value: true} optDtype := C.mlx_optional_dtype{has_value: false} + var globalScale C.mlx_array var b C.mlx_array if biases != nil { @@ -2161,7 +2163,7 @@ func Dequantize(w, scales, biases *Array, groupSize, bits int, mode string) *Arr } res := C.mlx_array_new() - C.mlx_dequantize(&res, w.c, scales.c, b, optGroupSize, optBits, cMode, optDtype, C.default_stream()) + C.mlx_dequantize(&res, w.c, scales.c, b, optGroupSize, optBits, cMode, globalScale, optDtype, C.default_stream()) return newArray(res) } diff --git a/x/imagegen/mlx/mlx.h b/x/imagegen/mlx/mlx.h index 34829d732..3f53c8941 100644 --- a/x/imagegen/mlx/mlx.h +++ b/x/imagegen/mlx/mlx.h @@ -309,10 +309,12 @@ #undef mlx_atleast_1d #undef mlx_atleast_2d #undef mlx_atleast_3d +#undef mlx_bartlett #undef mlx_bitwise_and #undef mlx_bitwise_invert #undef mlx_bitwise_or #undef mlx_bitwise_xor +#undef mlx_blackman #undef mlx_block_masked_mm #undef mlx_broadcast_arrays #undef mlx_broadcast_to @@ -365,6 +367,8 @@ #undef mlx_greater #undef mlx_greater_equal #undef mlx_hadamard_transform +#undef mlx_hamming +#undef mlx_hanning #undef mlx_identity #undef mlx_imag #undef mlx_inner @@ -751,8 +755,8 @@ extern int (*mlx_distributed_sum_scatter_ptr)(mlx_array* res, const mlx_array x, extern int (*mlx_distributed_group_rank_ptr)(mlx_distributed_group group); extern int (*mlx_distributed_group_size_ptr)(mlx_distributed_group group); extern mlx_distributed_group (*mlx_distributed_group_split_ptr)(mlx_distributed_group group, int color, int key); -extern bool (*mlx_distributed_is_available_ptr)(void); -extern mlx_distributed_group (*mlx_distributed_init_ptr)(bool strict); +extern bool (*mlx_distributed_is_available_ptr)(const char* bk); +extern mlx_distributed_group (*mlx_distributed_init_ptr)(bool strict, const char* bk); extern void (*mlx_set_error_handler_ptr)(mlx_error_handler_func handler, void* data, void (*dtor)(void*)); extern void (*_mlx_error_ptr)(const char* file, const int line, const char* fmt, ...); extern int (*mlx_export_function_ptr)(const char* file, const mlx_closure fun, const mlx_vector_array args, bool shapeless); @@ -905,10 +909,12 @@ extern int (*mlx_astype_ptr)(mlx_array* res, const mlx_array a, mlx_dtype dtype, extern int (*mlx_atleast_1d_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_atleast_2d_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_atleast_3d_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); +extern int (*mlx_bartlett_ptr)(mlx_array* res, int M, const mlx_stream s); extern int (*mlx_bitwise_and_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); extern int (*mlx_bitwise_invert_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_bitwise_or_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); extern int (*mlx_bitwise_xor_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); +extern int (*mlx_blackman_ptr)(mlx_array* res, int M, const mlx_stream s); extern int (*mlx_block_masked_mm_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, int block_size, const mlx_array mask_out , const mlx_array mask_lhs , const mlx_array mask_rhs , const mlx_stream s); extern int (*mlx_broadcast_arrays_ptr)(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_stream s); extern int (*mlx_broadcast_to_ptr)(mlx_array* res, const mlx_array a, const int* shape, size_t shape_num, const mlx_stream s); @@ -934,7 +940,7 @@ extern int (*mlx_cumprod_ptr)(mlx_array* res, const mlx_array a, int axis, bool extern int (*mlx_cumsum_ptr)(mlx_array* res, const mlx_array a, int axis, bool reverse, bool inclusive, const mlx_stream s); extern int (*mlx_degrees_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_depends_ptr)(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_vector_array dependencies); -extern int (*mlx_dequantize_ptr)(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, mlx_optional_dtype dtype, const mlx_stream s); +extern int (*mlx_dequantize_ptr)(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , mlx_optional_dtype dtype, const mlx_stream s); extern int (*mlx_diag_ptr)(mlx_array* res, const mlx_array a, int k, const mlx_stream s); extern int (*mlx_diagonal_ptr)(mlx_array* res, const mlx_array a, int offset, int axis1, int axis2, const mlx_stream s); extern int (*mlx_divide_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); @@ -961,6 +967,8 @@ extern int (*mlx_gather_qmm_ptr)(mlx_array* res, const mlx_array x, const mlx_ar extern int (*mlx_greater_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); extern int (*mlx_greater_equal_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); extern int (*mlx_hadamard_transform_ptr)(mlx_array* res, const mlx_array a, mlx_optional_float scale, const mlx_stream s); +extern int (*mlx_hamming_ptr)(mlx_array* res, int M, const mlx_stream s); +extern int (*mlx_hanning_ptr)(mlx_array* res, int M, const mlx_stream s); extern int (*mlx_identity_ptr)(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s); extern int (*mlx_imag_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_inner_ptr)(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); @@ -1020,8 +1028,8 @@ extern int (*mlx_prod_axes_ptr)(mlx_array* res, const mlx_array a, const int* ax extern int (*mlx_prod_axis_ptr)(mlx_array* res, const mlx_array a, int axis, bool keepdims, const mlx_stream s); extern int (*mlx_prod_ptr)(mlx_array* res, const mlx_array a, bool keepdims, const mlx_stream s); extern int (*mlx_put_along_axis_ptr)(mlx_array* res, const mlx_array a, const mlx_array indices, const mlx_array values, int axis, const mlx_stream s); -extern int (*mlx_qqmm_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s); -extern int (*mlx_quantize_ptr)(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s); +extern int (*mlx_qqmm_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale_x , const mlx_array global_scale_w , const mlx_stream s); +extern int (*mlx_quantize_ptr)(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , const mlx_stream s); extern int (*mlx_quantized_matmul_ptr)(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array scales, const mlx_array biases , bool transpose, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s); extern int (*mlx_radians_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_real_ptr)(mlx_array* res, const mlx_array a, const mlx_stream s); @@ -1492,9 +1500,9 @@ int mlx_distributed_group_size(mlx_distributed_group group); mlx_distributed_group mlx_distributed_group_split(mlx_distributed_group group, int color, int key); -bool mlx_distributed_is_available(void); +bool mlx_distributed_is_available(const char* bk); -mlx_distributed_group mlx_distributed_init(bool strict); +mlx_distributed_group mlx_distributed_init(bool strict, const char* bk); void mlx_set_error_handler(mlx_error_handler_func handler, void* data, void (*dtor)(void*)); @@ -1800,6 +1808,8 @@ int mlx_atleast_2d(mlx_array* res, const mlx_array a, const mlx_stream s); int mlx_atleast_3d(mlx_array* res, const mlx_array a, const mlx_stream s); +int mlx_bartlett(mlx_array* res, int M, const mlx_stream s); + int mlx_bitwise_and(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); int mlx_bitwise_invert(mlx_array* res, const mlx_array a, const mlx_stream s); @@ -1808,6 +1818,8 @@ int mlx_bitwise_or(mlx_array* res, const mlx_array a, const mlx_array b, const m int mlx_bitwise_xor(mlx_array* res, const mlx_array a, const mlx_array b, const mlx_stream s); +int mlx_blackman(mlx_array* res, int M, const mlx_stream s); + int mlx_block_masked_mm(mlx_array* res, const mlx_array a, const mlx_array b, int block_size, const mlx_array mask_out , const mlx_array mask_lhs , const mlx_array mask_rhs , const mlx_stream s); int mlx_broadcast_arrays(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_stream s); @@ -1858,7 +1870,7 @@ int mlx_degrees(mlx_array* res, const mlx_array a, const mlx_stream s); int mlx_depends(mlx_vector_array* res, const mlx_vector_array inputs, const mlx_vector_array dependencies); -int mlx_dequantize(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, mlx_optional_dtype dtype, const mlx_stream s); +int mlx_dequantize(mlx_array* res, const mlx_array w, const mlx_array scales, const mlx_array biases , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , mlx_optional_dtype dtype, const mlx_stream s); int mlx_diag(mlx_array* res, const mlx_array a, int k, const mlx_stream s); @@ -1912,6 +1924,10 @@ int mlx_greater_equal(mlx_array* res, const mlx_array a, const mlx_array b, cons int mlx_hadamard_transform(mlx_array* res, const mlx_array a, mlx_optional_float scale, const mlx_stream s); +int mlx_hamming(mlx_array* res, int M, const mlx_stream s); + +int mlx_hanning(mlx_array* res, int M, const mlx_stream s); + int mlx_identity(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s); int mlx_imag(mlx_array* res, const mlx_array a, const mlx_stream s); @@ -2030,9 +2046,9 @@ int mlx_prod(mlx_array* res, const mlx_array a, bool keepdims, const mlx_stream int mlx_put_along_axis(mlx_array* res, const mlx_array a, const mlx_array indices, const mlx_array values, int axis, const mlx_stream s); -int mlx_qqmm(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s); +int mlx_qqmm(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array w_scales , mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale_x , const mlx_array global_scale_w , const mlx_stream s); -int mlx_quantize(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s); +int mlx_quantize(mlx_vector_array* res, const mlx_array w, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_array global_scale , const mlx_stream s); int mlx_quantized_matmul(mlx_array* res, const mlx_array x, const mlx_array w, const mlx_array scales, const mlx_array biases , bool transpose, mlx_optional_int group_size, mlx_optional_int bits, const char* mode, const mlx_stream s); diff --git a/x/mlxrunner/mlx/CMakeLists.txt b/x/mlxrunner/mlx/CMakeLists.txt index 9825c441b..9ba6f62e3 100644 --- a/x/mlxrunner/mlx/CMakeLists.txt +++ b/x/mlxrunner/mlx/CMakeLists.txt @@ -15,7 +15,8 @@ set(CMAKE_INSTALL_RPATH "@loader_path") include(FetchContent) -set(MLX_C_GIT_TAG "v0.5.0" CACHE STRING "") +file(READ "${CMAKE_CURRENT_SOURCE_DIR}/../../../MLX_VERSION" MLX_C_GIT_TAG) +string(STRIP "${MLX_C_GIT_TAG}" MLX_C_GIT_TAG) FetchContent_Declare( mlx-c diff --git a/x/mlxrunner/mlx/generated.c b/x/mlxrunner/mlx/generated.c index ecf9d30c8..f1333971c 100644 --- a/x/mlxrunner/mlx/generated.c +++ b/x/mlxrunner/mlx/generated.c @@ -326,8 +326,10 @@ int (*mlx_distributed_sum_scatter_)( int (*mlx_distributed_group_rank_)(mlx_distributed_group group) = NULL; int (*mlx_distributed_group_size_)(mlx_distributed_group group) = NULL; mlx_distributed_group (*mlx_distributed_group_split_)(mlx_distributed_group group, int color, int key) = NULL; -bool (*mlx_distributed_is_available_)(void) = NULL; -mlx_distributed_group (*mlx_distributed_init_)(bool strict) = NULL; +bool (*mlx_distributed_is_available_)(const char* bk /* may be null */) = NULL; +mlx_distributed_group (*mlx_distributed_init_)( + bool strict, + const char* bk /* may be null */) = NULL; void (*mlx_set_error_handler_)( mlx_error_handler_func handler, void* data, @@ -924,6 +926,7 @@ int (*mlx_astype_)( int (*mlx_atleast_1d_)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_atleast_2d_)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_atleast_3d_)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; +int (*mlx_bartlett_)(mlx_array* res, int M, const mlx_stream s) = NULL; int (*mlx_bitwise_and_)( mlx_array* res, const mlx_array a, @@ -940,6 +943,7 @@ int (*mlx_bitwise_xor_)( const mlx_array a, const mlx_array b, const mlx_stream s) = NULL; +int (*mlx_blackman_)(mlx_array* res, int M, const mlx_stream s) = NULL; int (*mlx_block_masked_mm_)( mlx_array* res, const mlx_array a, @@ -1120,6 +1124,7 @@ int (*mlx_dequantize_)( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, mlx_optional_dtype dtype, const mlx_stream s) = NULL; int (*mlx_diag_)(mlx_array* res, const mlx_array a, int k, const mlx_stream s) = NULL; @@ -1256,6 +1261,8 @@ int (*mlx_hadamard_transform_)( const mlx_array a, mlx_optional_float scale, const mlx_stream s) = NULL; +int (*mlx_hamming_)(mlx_array* res, int M, const mlx_stream s) = NULL; +int (*mlx_hanning_)(mlx_array* res, int M, const mlx_stream s) = NULL; int (*mlx_identity_)(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s) = NULL; int (*mlx_imag_)(mlx_array* res, const mlx_array a, const mlx_stream s) = NULL; int (*mlx_inner_)( @@ -1548,6 +1555,8 @@ int (*mlx_qqmm_)( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale_x /* may be null */, + const mlx_array global_scale_w /* may be null */, const mlx_stream s) = NULL; int (*mlx_quantize_)( mlx_vector_array* res, @@ -1555,6 +1564,7 @@ int (*mlx_quantize_)( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, const mlx_stream s) = NULL; int (*mlx_quantized_matmul_)( mlx_array* res, @@ -2550,10 +2560,12 @@ int mlx_dynamic_load_symbols(mlx_dynamic_handle handle) { CHECK_LOAD(handle, mlx_atleast_1d); CHECK_LOAD(handle, mlx_atleast_2d); CHECK_LOAD(handle, mlx_atleast_3d); + CHECK_LOAD(handle, mlx_bartlett); CHECK_LOAD(handle, mlx_bitwise_and); CHECK_LOAD(handle, mlx_bitwise_invert); CHECK_LOAD(handle, mlx_bitwise_or); CHECK_LOAD(handle, mlx_bitwise_xor); + CHECK_LOAD(handle, mlx_blackman); CHECK_LOAD(handle, mlx_block_masked_mm); CHECK_LOAD(handle, mlx_broadcast_arrays); CHECK_LOAD(handle, mlx_broadcast_to); @@ -2606,6 +2618,8 @@ int mlx_dynamic_load_symbols(mlx_dynamic_handle handle) { CHECK_LOAD(handle, mlx_greater); CHECK_LOAD(handle, mlx_greater_equal); CHECK_LOAD(handle, mlx_hadamard_transform); + CHECK_LOAD(handle, mlx_hamming); + CHECK_LOAD(handle, mlx_hanning); CHECK_LOAD(handle, mlx_identity); CHECK_LOAD(handle, mlx_imag); CHECK_LOAD(handle, mlx_inner); diff --git a/x/mlxrunner/mlx/generated.h b/x/mlxrunner/mlx/generated.h index e8dfa7b90..26119f2ff 100644 --- a/x/mlxrunner/mlx/generated.h +++ b/x/mlxrunner/mlx/generated.h @@ -300,10 +300,12 @@ #define mlx_atleast_1d mlx_atleast_1d_mlx_gen_orig_ #define mlx_atleast_2d mlx_atleast_2d_mlx_gen_orig_ #define mlx_atleast_3d mlx_atleast_3d_mlx_gen_orig_ +#define mlx_bartlett mlx_bartlett_mlx_gen_orig_ #define mlx_bitwise_and mlx_bitwise_and_mlx_gen_orig_ #define mlx_bitwise_invert mlx_bitwise_invert_mlx_gen_orig_ #define mlx_bitwise_or mlx_bitwise_or_mlx_gen_orig_ #define mlx_bitwise_xor mlx_bitwise_xor_mlx_gen_orig_ +#define mlx_blackman mlx_blackman_mlx_gen_orig_ #define mlx_block_masked_mm mlx_block_masked_mm_mlx_gen_orig_ #define mlx_broadcast_arrays mlx_broadcast_arrays_mlx_gen_orig_ #define mlx_broadcast_to mlx_broadcast_to_mlx_gen_orig_ @@ -356,6 +358,8 @@ #define mlx_greater mlx_greater_mlx_gen_orig_ #define mlx_greater_equal mlx_greater_equal_mlx_gen_orig_ #define mlx_hadamard_transform mlx_hadamard_transform_mlx_gen_orig_ +#define mlx_hamming mlx_hamming_mlx_gen_orig_ +#define mlx_hanning mlx_hanning_mlx_gen_orig_ #define mlx_identity mlx_identity_mlx_gen_orig_ #define mlx_imag mlx_imag_mlx_gen_orig_ #define mlx_inner mlx_inner_mlx_gen_orig_ @@ -889,10 +893,12 @@ #undef mlx_atleast_1d #undef mlx_atleast_2d #undef mlx_atleast_3d +#undef mlx_bartlett #undef mlx_bitwise_and #undef mlx_bitwise_invert #undef mlx_bitwise_or #undef mlx_bitwise_xor +#undef mlx_blackman #undef mlx_block_masked_mm #undef mlx_broadcast_arrays #undef mlx_broadcast_to @@ -945,6 +951,8 @@ #undef mlx_greater #undef mlx_greater_equal #undef mlx_hadamard_transform +#undef mlx_hamming +#undef mlx_hanning #undef mlx_identity #undef mlx_imag #undef mlx_inner @@ -1501,8 +1509,10 @@ extern int (*mlx_distributed_sum_scatter_)( extern int (*mlx_distributed_group_rank_)(mlx_distributed_group group); extern int (*mlx_distributed_group_size_)(mlx_distributed_group group); extern mlx_distributed_group (*mlx_distributed_group_split_)(mlx_distributed_group group, int color, int key); -extern bool (*mlx_distributed_is_available_)(void); -extern mlx_distributed_group (*mlx_distributed_init_)(bool strict); +extern bool (*mlx_distributed_is_available_)(const char* bk /* may be null */); +extern mlx_distributed_group (*mlx_distributed_init_)( + bool strict, + const char* bk /* may be null */); extern void (*mlx_set_error_handler_)( mlx_error_handler_func handler, void* data, @@ -2099,6 +2109,7 @@ extern int (*mlx_astype_)( extern int (*mlx_atleast_1d_)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_atleast_2d_)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_atleast_3d_)(mlx_array* res, const mlx_array a, const mlx_stream s); +extern int (*mlx_bartlett_)(mlx_array* res, int M, const mlx_stream s); extern int (*mlx_bitwise_and_)( mlx_array* res, const mlx_array a, @@ -2115,6 +2126,7 @@ extern int (*mlx_bitwise_xor_)( const mlx_array a, const mlx_array b, const mlx_stream s); +extern int (*mlx_blackman_)(mlx_array* res, int M, const mlx_stream s); extern int (*mlx_block_masked_mm_)( mlx_array* res, const mlx_array a, @@ -2295,6 +2307,7 @@ extern int (*mlx_dequantize_)( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, mlx_optional_dtype dtype, const mlx_stream s); extern int (*mlx_diag_)(mlx_array* res, const mlx_array a, int k, const mlx_stream s); @@ -2431,6 +2444,8 @@ extern int (*mlx_hadamard_transform_)( const mlx_array a, mlx_optional_float scale, const mlx_stream s); +extern int (*mlx_hamming_)(mlx_array* res, int M, const mlx_stream s); +extern int (*mlx_hanning_)(mlx_array* res, int M, const mlx_stream s); extern int (*mlx_identity_)(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s); extern int (*mlx_imag_)(mlx_array* res, const mlx_array a, const mlx_stream s); extern int (*mlx_inner_)( @@ -2723,6 +2738,8 @@ extern int (*mlx_qqmm_)( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale_x /* may be null */, + const mlx_array global_scale_w /* may be null */, const mlx_stream s); extern int (*mlx_quantize_)( mlx_vector_array* res, @@ -2730,6 +2747,7 @@ extern int (*mlx_quantize_)( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, const mlx_stream s); extern int (*mlx_quantized_matmul_)( mlx_array* res, @@ -4033,11 +4051,13 @@ static inline int mlx_distributed_group_size(mlx_distributed_group group) { static inline mlx_distributed_group mlx_distributed_group_split(mlx_distributed_group group, int color, int key) { return mlx_distributed_group_split_(group, color, key); } -static inline bool mlx_distributed_is_available(void) { - return mlx_distributed_is_available_(); +static inline bool mlx_distributed_is_available(const char* bk /* may be null */) { + return mlx_distributed_is_available_(bk); } -static inline mlx_distributed_group mlx_distributed_init(bool strict) { - return mlx_distributed_init_(strict); +static inline mlx_distributed_group mlx_distributed_init( + bool strict, + const char* bk /* may be null */) { + return mlx_distributed_init_(strict, bk); } static inline void mlx_set_error_handler( mlx_error_handler_func handler, @@ -4939,6 +4959,9 @@ static inline int mlx_atleast_2d(mlx_array* res, const mlx_array a, const mlx_st static inline int mlx_atleast_3d(mlx_array* res, const mlx_array a, const mlx_stream s) { return mlx_atleast_3d_(res, a, s); } +static inline int mlx_bartlett(mlx_array* res, int M, const mlx_stream s) { + return mlx_bartlett_(res, M, s); +} static inline int mlx_bitwise_and( mlx_array* res, const mlx_array a, @@ -4963,6 +4986,9 @@ static inline int mlx_bitwise_xor( const mlx_stream s) { return mlx_bitwise_xor_(res, a, b, s); } +static inline int mlx_blackman(mlx_array* res, int M, const mlx_stream s) { + return mlx_blackman_(res, M, s); +} static inline int mlx_block_masked_mm( mlx_array* res, const mlx_array a, @@ -5193,9 +5219,10 @@ static inline int mlx_dequantize( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, mlx_optional_dtype dtype, const mlx_stream s) { - return mlx_dequantize_(res, w, scales, biases, group_size, bits, mode, dtype, s); + return mlx_dequantize_(res, w, scales, biases, group_size, bits, mode, global_scale, dtype, s); } static inline int mlx_diag(mlx_array* res, const mlx_array a, int k, const mlx_stream s) { return mlx_diag_(res, a, k, s); @@ -5383,6 +5410,12 @@ static inline int mlx_hadamard_transform( const mlx_stream s) { return mlx_hadamard_transform_(res, a, scale, s); } +static inline int mlx_hamming(mlx_array* res, int M, const mlx_stream s) { + return mlx_hamming_(res, M, s); +} +static inline int mlx_hanning(mlx_array* res, int M, const mlx_stream s) { + return mlx_hanning_(res, M, s); +} static inline int mlx_identity(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s) { return mlx_identity_(res, n, dtype, s); } @@ -5793,8 +5826,10 @@ static inline int mlx_qqmm( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale_x /* may be null */, + const mlx_array global_scale_w /* may be null */, const mlx_stream s) { - return mlx_qqmm_(res, x, w, w_scales, group_size, bits, mode, s); + return mlx_qqmm_(res, x, w, w_scales, group_size, bits, mode, global_scale_x, global_scale_w, s); } static inline int mlx_quantize( mlx_vector_array* res, @@ -5802,8 +5837,9 @@ static inline int mlx_quantize( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, const mlx_stream s) { - return mlx_quantize_(res, w, group_size, bits, mode, s); + return mlx_quantize_(res, w, group_size, bits, mode, global_scale, s); } static inline int mlx_quantized_matmul( mlx_array* res, diff --git a/x/mlxrunner/mlx/include/mlx/c/distributed_group.h b/x/mlxrunner/mlx/include/mlx/c/distributed_group.h index 3cfccc806..43aa2ae56 100644 --- a/x/mlxrunner/mlx/include/mlx/c/distributed_group.h +++ b/x/mlxrunner/mlx/include/mlx/c/distributed_group.h @@ -42,12 +42,14 @@ mlx_distributed_group_split(mlx_distributed_group group, int color, int key); /** * Check if distributed is available. */ -bool mlx_distributed_is_available(void); +bool mlx_distributed_is_available(const char* bk /* may be null */); /** * Initialize distributed. */ -mlx_distributed_group mlx_distributed_init(bool strict); +mlx_distributed_group mlx_distributed_init( + bool strict, + const char* bk /* may be null */); /**@}*/ diff --git a/x/mlxrunner/mlx/include/mlx/c/ops.h b/x/mlxrunner/mlx/include/mlx/c/ops.h index a1446fb9e..64d70e2f4 100644 --- a/x/mlxrunner/mlx/include/mlx/c/ops.h +++ b/x/mlxrunner/mlx/include/mlx/c/ops.h @@ -166,6 +166,7 @@ int mlx_astype( int mlx_atleast_1d(mlx_array* res, const mlx_array a, const mlx_stream s); int mlx_atleast_2d(mlx_array* res, const mlx_array a, const mlx_stream s); int mlx_atleast_3d(mlx_array* res, const mlx_array a, const mlx_stream s); +int mlx_bartlett(mlx_array* res, int M, const mlx_stream s); int mlx_bitwise_and( mlx_array* res, const mlx_array a, @@ -182,6 +183,7 @@ int mlx_bitwise_xor( const mlx_array a, const mlx_array b, const mlx_stream s); +int mlx_blackman(mlx_array* res, int M, const mlx_stream s); int mlx_block_masked_mm( mlx_array* res, const mlx_array a, @@ -362,6 +364,7 @@ int mlx_dequantize( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, mlx_optional_dtype dtype, const mlx_stream s); int mlx_diag(mlx_array* res, const mlx_array a, int k, const mlx_stream s); @@ -498,6 +501,8 @@ int mlx_hadamard_transform( const mlx_array a, mlx_optional_float scale, const mlx_stream s); +int mlx_hamming(mlx_array* res, int M, const mlx_stream s); +int mlx_hanning(mlx_array* res, int M, const mlx_stream s); int mlx_identity(mlx_array* res, int n, mlx_dtype dtype, const mlx_stream s); int mlx_imag(mlx_array* res, const mlx_array a, const mlx_stream s); int mlx_inner( @@ -790,6 +795,8 @@ int mlx_qqmm( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale_x /* may be null */, + const mlx_array global_scale_w /* may be null */, const mlx_stream s); int mlx_quantize( mlx_vector_array* res, @@ -797,6 +804,7 @@ int mlx_quantize( mlx_optional_int group_size, mlx_optional_int bits, const char* mode, + const mlx_array global_scale /* may be null */, const mlx_stream s); int mlx_quantized_matmul( mlx_array* res, diff --git a/x/mlxrunner/mlx/ops_extra.go b/x/mlxrunner/mlx/ops_extra.go index ff06092e9..ee9bbf6e5 100644 --- a/x/mlxrunner/mlx/ops_extra.go +++ b/x/mlxrunner/mlx/ops_extra.go @@ -15,9 +15,10 @@ func Quantize(w *Array, groupSize, bits int, mode string) (weights, scales, bias defer C.free(unsafe.Pointer(cMode)) optGroupSize := C.mlx_optional_int{value: C.int(groupSize), has_value: true} optBits := C.mlx_optional_int{value: C.int(bits), has_value: true} + var globalScale C.mlx_array res := C.mlx_vector_array_new() defer C.mlx_vector_array_free(res) - C.mlx_quantize(&res, w.ctx, optGroupSize, optBits, cMode, DefaultStream().ctx) + C.mlx_quantize(&res, w.ctx, optGroupSize, optBits, cMode, globalScale, DefaultStream().ctx) vecSize := int(C.mlx_vector_array_size(res)) w0 := New("QUANTIZE_W") @@ -38,6 +39,7 @@ func Dequantize(w, scales, biases *Array, groupSize, bits int, mode string) *Arr optGroupSize := C.mlx_optional_int{value: C.int(groupSize), has_value: true} optBits := C.mlx_optional_int{value: C.int(bits), has_value: true} optDtype := C.mlx_optional_dtype{has_value: false} + var globalScale C.mlx_array var b C.mlx_array if biases != nil { @@ -45,7 +47,7 @@ func Dequantize(w, scales, biases *Array, groupSize, bits int, mode string) *Arr } out := New("DEQUANTIZE") - C.mlx_dequantize(&out.ctx, w.ctx, scales.ctx, b, optGroupSize, optBits, cMode, optDtype, DefaultStream().ctx) + C.mlx_dequantize(&out.ctx, w.ctx, scales.ctx, b, optGroupSize, optBits, cMode, globalScale, optDtype, DefaultStream().ctx) return out }