whispercpp 1.3.3 → 1.3.4

data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh

@@ -1,8 +1,20 @@
 #pragma once
 
 #include "common.cuh"
+
 #include <cstdint>
 
+static __device__ __forceinline__ int get_int_b1(const void * x, const int & i32) {
+    const uint8_t * x8 = (const uint8_t *) x;
+
+    int x32  = x8[4*i32 + 0] <<  0;
+    x32     |= x8[4*i32 + 1] <<  8;
+    x32     |= x8[4*i32 + 2] << 16;
+    x32     |= x8[4*i32 + 3] << 24;
+
+    return x32;
+}
+
 static __device__ __forceinline__ int get_int_b2(const void * x, const int & i32) {
     const uint16_t * x16 = (const uint16_t *) x; // assume at least 2 byte alignment
 
@@ -16,6 +28,72 @@ static __device__ __forceinline__ int get_int_b4(const void * x, const int & i32
     return ((const int *) x)[i32]; // assume at least 4 byte alignment
 }
 
+// q4 contains 8 indices with 4 bit each.
+// This function selects those bytes from table that are at those indices and returns them as int2.
+// The first int contains the bytes with even indices in q4, the second int contains the bytes with odd indices in q4.
+static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, const int8_t * table) {
+#if defined(GGML_USE_HIP)
+    // Load the 16-byte table into four 32-bit unsigned integers.
+    const uint32_t *values = (const uint32_t *)table;
+
+    const uint32_t q_even = q4;
+    const uint32_t q_odd  = (q4 >> 4);
+
+    // Perform lookups in the lower half of the table (indices 0-7).
+    uint32_t v_even_low = __builtin_amdgcn_perm(values[1], values[0], q_even & 0x07070707);
+    uint32_t v_odd_low = __builtin_amdgcn_perm(values[1], values[0], q_odd & 0x07070707);
+
+    // Perform lookups in the upper half of the table (indices 8-15).
+    uint32_t v_even_high = __builtin_amdgcn_perm(values[3], values[2], q_even & 0x07070707);
+    uint32_t v_odd_high = __builtin_amdgcn_perm(values[3], values[2], q_odd & 0x07070707);
+
+    // Select between the low and high results based on the MSB of each index nibble.
+    uint32_t mask_even = 0x03020100 | ((q_even & 0x08080808) >> 1);
+    uint32_t res_x = __builtin_amdgcn_perm(v_even_high, v_even_low, mask_even);
+    uint32_t mask_odd = 0x03020100 | ((q_odd & 0x08080808) >> 1);
+    uint32_t res_y = __builtin_amdgcn_perm(v_odd_high, v_odd_low, mask_odd);
+
+    return make_int2(res_x, res_y);
+#elif !defined(GGML_USE_MUSA)
+    // CUDA does not have an instruction for selecting bytes with 4 bit indices.
+    // However, __byte_perm is an instruction that selects bytes with 3 bit indices that can be used instead.
+    const uint32_t * table32 = (const uint32_t *) table;
+
+    // __byte_perm selects bytes based on the lower 16 bits in its third argument.
+    // Therefore, do 2 iterations over the 32 bits in q4 with 0 and 16 shift.
+    // To handle the fourth bit, first call _byte_perm both for the low and the high 64 bit of table, using the low 3 bits.
+    // Then, call __byte_perm again to select from the low and high bytes based on the fourth bit.
+    uint32_t tmp[2];
+    const uint32_t low_high_selection_indices = (0x32103210 | ((q4 & 0x88888888) >> 1));
+#pragma unroll
+    for (uint32_t i = 0; i < 2; ++i) {
+        const uint32_t shift = 16 * i;
+
+        const uint32_t low  = __byte_perm(table32[0], table32[1], q4 >> shift);
+        const uint32_t high = __byte_perm(table32[2], table32[3], q4 >> shift);
+        tmp[i] = __byte_perm(low, high, low_high_selection_indices >> shift);
+    }
+
+    // tmp contains the bytes from tyble in the same order as the 4 bit indices in q4.
+    // However, for the result we need ints with all even/odd 4 bit indices in q4.
+    // Therefore, 2 more calls to __byte_perm to put the bytes in the correct order.
+    return make_int2(__byte_perm(tmp[0], tmp[1], 0x6420), __byte_perm(tmp[0], tmp[1], 0x7531));
+#else
+    // Generic implementation.
+    const int      q0_32  = (q4 >> 0) & 0x0F0F0F0F;
+    const int8_t * q0_8   = (const int8_t *) &q0_32;
+    const char4    val0_8 = make_char4(
+        table[q0_8[0]], table[q0_8[1]], table[q0_8[2]], table[q0_8[3]]);
+
+    const int      q1_32  = (q4 >> 4) & 0x0F0F0F0F;
+    const int8_t * q1_8   = (const int8_t *) &q1_32;
+    const char4    val1_8 = make_char4(
+        table[q1_8[0]], table[q1_8[1]], table[q1_8[2]], table[q1_8[3]]);
+
+    return make_int2(*((const int *) &val0_8), *((const int *) &val1_8));
+#endif
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
@@ -61,7 +139,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp
         sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi);
     }
 
-#ifdef GGML_CUDA_F16
+#ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(__hmul2(dm4, ds8));
     const float d4d8 = tmp.x;
     const float m4s8 = tmp.y;
@@ -70,7 +148,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp
     const float2 ds8f = __half22float2(ds8);
     const float d4d8 = dm4f.x * ds8f.x;
     const float m4s8 = dm4f.y * ds8f.y;
-#endif // GGML_CUDA_F16
+#endif // FAST_FP16_AVAILABLE
 
     // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
     return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
@@ -132,7 +210,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp
         sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
     }
 
-#ifdef GGML_CUDA_F16
+#ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(__hmul2(dm5, ds8));
     const float d5d8 = tmp.x;
     const float m5s8 = tmp.y;
@@ -141,7 +219,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp
     const float2 ds8f = __half22float2(ds8);
     const float d5d8 = dm5f.x * ds8f.x;
     const float m5s8 = dm5f.y * ds8f.y;
-#endif // GGML_CUDA_F16
+#endif // FAST_FP16_AVAILABLE
 
     // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
     return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
@@ -175,7 +253,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
         sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
     }
 
-#ifdef GGML_CUDA_F16
+#ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(__hmul2(dm8, ds8));
     const float d8d8 = tmp.x;
     const float m8s8 = tmp.y;
@@ -184,7 +262,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
     const float2 ds8f = __half22float2(ds8);
     const float d8d8 = dm8f.x * ds8f.x;
     const float m8s8 = dm8f.y * ds8f.y;
-#endif // GGML_CUDA_F16
+#endif // FAST_FP16_AVAILABLE
 
     // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
     return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
@@ -211,6 +289,30 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_16_q8_1_
     return d8_1*sumf;
 }
 
+#define VDR_MXFP4_Q8_1_MMVQ 2
+#define VDR_MXFP4_Q8_1_MMQ  4
+
+static __device__ __forceinline__ float vec_dot_mxfp4_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_mxfp4 * bq4 = (const block_mxfp4 *) vbq + kbx;
+
+    const int * q8 = (const int *) bq8_1->qs + iqs;
+
+    int sumi = 0;
+#pragma unroll
+    for (int l = 0; l < VDR_MXFP4_Q8_1_MMVQ; ++l) {
+        const int aux_q4 = get_int_b1(bq4->qs, iqs + l);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4);
+
+        sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi);
+        sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi);
+    }
+
+    const float d = ggml_cuda_e8m0_to_fp32(bq4->e) * 0.5f * __low2float(bq8_1->ds);
+    return d * sumi;
+}
+
 #define VDR_Q2_K_Q8_1_MMVQ 1
 #define VDR_Q2_K_Q8_1_MMQ  4
 
@@ -1068,20 +1170,6 @@ static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
     return d * ((sumi[0] + sumf[0]) * sc0 + (sumi[1] + sumf[1]) * sc1);
 }
 
-static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4) {
-    const int      q0_32  = (q4 >> 0) & 0x0F0F0F0F;
-    const int8_t * q0_8   = (const int8_t *) &q0_32;
-    const char4    val0_8 = make_char4(
-        kvalues_iq4nl[q0_8[0]], kvalues_iq4nl[q0_8[1]], kvalues_iq4nl[q0_8[2]], kvalues_iq4nl[q0_8[3]]);
-
-    const int      q1_32  = (q4 >> 4) & 0x0F0F0F0F;
-    const int8_t * q1_8   = (const int8_t *) &q1_32;
-    const char4    val1_8 = make_char4(
-        kvalues_iq4nl[q1_8[0]], kvalues_iq4nl[q1_8[1]], kvalues_iq4nl[q1_8[2]], kvalues_iq4nl[q1_8[3]]);
-
-    return make_int2(*((const int *) &val0_8), *((const int *) &val1_8));
-}
-
 #define VDR_IQ4_NL_Q8_1_MMVQ 2
 #define VDR_IQ4_NL_Q8_1_MMQ  4
 
@@ -1096,7 +1184,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
 #pragma unroll
     for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
         const int aux_q4 = get_int_b2(bq4->qs, iqs + l);
-        const int2 v = get_int_from_table_16(aux_q4);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
 
         sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi);
         sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi);
@@ -1118,7 +1206,7 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
 #pragma unroll
     for (int j = 0; j < 4; ++j) {
         const int aux_q4 = get_int_b4(bq4->qs, iqs + j);
-        const int2 v = get_int_from_table_16(aux_q4);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
 
         const int u0 = get_int_b4(bq8_1[iqs/4].qs, j + 0);
         const int u1 = get_int_b4(bq8_1[iqs/4].qs, j + 4);

data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h

@@ -6,6 +6,10 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 
+#if CUDART_VERSION >= 12050
+#include <cuda_fp8.h>
+#endif // CUDART_VERSION >= 12050
+
 #if CUDART_VERSION < 11020
 #define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
 #define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH

data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h

@@ -1,18 +1,11 @@
 #pragma once
 
-#define HIP_ENABLE_WARP_SYNC_BUILTINS 1
+#define HIP_DISABLE_WARP_SYNC_BUILTINS 1
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
-#include <hip/hip_bfloat16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
+#include <hip/hip_bf16.h>
 
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N
@@ -29,8 +22,10 @@
 #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
+#define __shfl_up_sync(mask, var, laneMask, width) __shfl_up(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define __all_sync(mask, var) __all(var)
+#define __any_sync(mask, var) __any(var)
 #define cublasCreate hipblasCreate
 #define cublasDestroy hipblasDestroy
 #define cublasGemmEx hipblasGemmEx
@@ -42,7 +37,6 @@
 #define cublasSgemm hipblasSgemm
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
 #define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
@@ -144,24 +138,61 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
+#if HIP_VERSION >= 60500000
+#define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
+#define cublasComputeType_t hipblasComputeType_t
+#define cudaDataType_t hipDataType
+#else
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define cublasComputeType_t hipblasDatatype_t
+#define cudaDataType_t hipblasDatatype_t
+#endif // HIP_VERSION >= 6050000
+
+#if !defined(__HIP_PLATFORM_AMD__)
+#error "The HIP backend supports only AMD targets"
+#endif // !defined(__HIP_PLATFORM_AMD__)
+
 #define __CUDA_ARCH__ 1300
 
-#if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)
+#if defined(__gfx900__) || defined(__gfx906__)
+#define GCN5
+#endif // defined(__gfx900__) || defined(__gfx906__)
+
+#if defined(__gfx803__)
+#define GCN4
+#endif // defined(__gfx803__)
+
+#if defined(GCN5) || defined(GCN4)
 #define GCN
-#endif
+#endif // defined(GCN5) || defined(GCN4)
 
-#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__)
-#define CDNA
-#endif
+#if defined(__gfx942__)
+#define CDNA3
+#endif // defined(__gfx942__)
+
+#if defined(__gfx90a__)
+#define CDNA2
+#endif // defined(__gfx90a__)
+
+#if defined(__gfx908__)
+#define CDNA1
+#endif // defined(__gfx908__)
+
+#if defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
+#define CDNA // For the entire family
+#endif // defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
 
 #if defined(__GFX12__)
 #define RDNA4
-#endif
+#endif // defined(__GFX12__)
 
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
+#if defined(__GFX11__)
 #define RDNA3
-#endif
+#endif // defined(__GFX11__)
 
 #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
     defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
@@ -170,13 +201,18 @@
 
 #if defined(__gfx1010__) || defined(__gfx1012__)
 #define RDNA1
-#endif
+#endif // defined(__gfx1010__) || defined(__gfx1012__)
+
+#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(RDNA1)
+#define RDNA // For the entire family
+#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(RDNA1)
 
 #ifndef __has_builtin
     #define __has_builtin(x) 0
 #endif
 
-typedef hip_bfloat16 nv_bfloat16;
+typedef __hip_bfloat16 nv_bfloat16;
+typedef __hip_bfloat162 nv_bfloat162;
 
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
@@ -227,17 +263,3 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne
     }
     return c;
 }
-
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
-    typedef union half2_b32 {
-        half2 val;
-        int   b32;
-    } half2_b32_t;
-    half2_b32_t tmp;
-    tmp.val = var;
-    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
-    return tmp.val;
-}
-#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000

data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h

@@ -13,7 +13,7 @@
 #define CUBLAS_OP_N MUBLAS_OP_N
 #define CUBLAS_OP_T MUBLAS_OP_T
 #define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
-#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_TENSOR_OP_MATH
 #define CUDA_R_16F  MUSA_R_16F
 #define CUDA_R_16BF MUSA_R_16BF
 #define CUDA_R_32F  MUSA_R_32F
@@ -29,7 +29,7 @@
 #define cublasSgemm mublasSgemm
 #define cublasStatus_t mublasStatus_t
 #define cublasOperation_t mublasOperation_t
-#define cublasGetStatusString mublasStatus_to_string
+#define cublasGetStatusString mublasGetStatusString
 #define cudaDataType_t musaDataType_t
 #define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
@@ -137,4 +137,5 @@
 #define cudaStreamEndCapture musaStreamEndCapture
 #define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
 
-typedef mt_bfloat16 nv_bfloat16;
+typedef __mt_bfloat16 nv_bfloat16;
+typedef __mt_bfloat162 nv_bfloat162;

data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt

@@ -46,8 +46,8 @@ if (GGML_HIP_ROCWMMA_FATTN)
     endif()
 endif()
 
-if (${hip_VERSION} VERSION_LESS 5.5)
-    message(FATAL_ERROR "At least ROCM/HIP V5.5 is required")
+if (${hip_VERSION} VERSION_LESS 6.1)
+    message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
 endif()
 
 message(STATUS "HIP and hipBLAS found")
@@ -113,10 +113,18 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (NOT GGML_HIP_MMQ_MFMA)
+    add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
+endif()
+
 if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
 endif()
 
+if (GGML_HIP_EXPORT_METRICS)
+    set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
+endif()
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()

data/ext/sources/ggml/src/ggml-impl.h

@@ -73,6 +73,35 @@ static inline int ggml_up(int n, int m) {
     return (n + m - 1) & ~(m - 1);
 }
 
+// TODO: move to ggml.h? (won't be able to inline)
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool ggml_op_is_empty(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return true;
+        default:
+            return false;
+    }
+}
+
 //
 // logging
 //
@@ -313,6 +342,10 @@ struct ggml_cgraph {
 // if you need the gradients, get them from the original graph
 struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 
+// ggml-alloc.c: true if the operation can reuse memory from its sources
+GGML_API bool ggml_op_can_inplace(enum ggml_op op);
+
+
 // Memory allocation
 
 GGML_API void * ggml_aligned_malloc(size_t size);
@@ -394,6 +427,67 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 
+static inline float ggml_e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;  // Stores the raw bit representation of the float
+
+    // Handle special case for minimum exponent (denormalized float)
+    if (x == 0) {
+        // Bit pattern for 2^(-127):
+        // - Sign bit: 0 (positive)
+        // - Exponent: 0 (denormalized number)
+        // - Mantissa: 0x400000 (0.5 in fractional form)
+        // Value = 0.5 * 2^(-126) = 2^(-127)
+        bits = 0x00400000;
+    }
+    // note: disabled as we don't need to handle NaNs
+    //// Handle special case for NaN (all bits set)
+    //else if (x == 0xFF) {
+    //    // Standard quiet NaN pattern:
+    //    // - Sign bit: 0
+    //    // - Exponent: all 1s (0xFF)
+    //    // - Mantissa: 0x400000 (quiet NaN flag)
+    //    bits = 0x7FC00000;
+    //}
+    // Normalized values (most common case)
+    else {
+        // Construct normalized float by shifting exponent into position:
+        // - Exponent field: 8 bits (positions 30-23)
+        // - Mantissa: 0 (implicit leading 1)
+        // Value = 2^(x - 127)
+        bits = (uint32_t) x << 23;
+    }
+
+    float result;  // Final float value
+                   // Safely reinterpret bit pattern as float without type-punning issues
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
+// Equal to ggml_e8m0_to_fp32/2
+// Useful with MXFP4 quantization since the E0M2 values are doubled
+static inline float ggml_e8m0_to_fp32_half(uint8_t x) {
+    uint32_t bits;
+
+    // For x < 2: use precomputed denormal patterns
+    if (x < 2) {
+        // 0x00200000 = 2^(-128), 0x00400000 = 2^(-127)
+        bits = 0x00200000 << x;
+    }
+    // For x >= 2: normalized exponent adjustment
+    else {
+        // 0.5 * 2^(x-127) = 2^(x-128) = normalized with exponent (x-1)
+        bits = (uint32_t)(x - 1) << 23;
+    }
+    // Note: NaNs are not handled here
+
+    float result;
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
+#define GGML_E8M0_TO_FP32(x) ggml_e8m0_to_fp32(x)
+#define GGML_E8M0_TO_FP32_HALF(x) ggml_e8m0_to_fp32_half(x)
+
 /**
  * Converts brain16 to float32.
  *
@@ -493,27 +587,27 @@ static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int n
     return true;
 }
 
-// Returns true if nodes [i, i+ops.size()) are the sequence of ggml_ops in ops[]
+// Returns true if nodes with indices { node_idxs } are the sequence of ggml_ops in ops[]
 // and are fusable. Nodes are considered fusable according to this function if:
 // - all nodes except the last have only one use and are not views/outputs (see ggml_node_has_N_uses).
 // - all nodes except the last are a src of the following node.
 // - all nodes are the same shape.
 // TODO: Consider allowing GGML_OP_NONE nodes in between
-static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, const enum ggml_op * ops, int num_ops) {
-    if (node_idx + num_ops > cgraph->n_nodes) {
-        return false;
-    }
-
+static inline bool ggml_can_fuse_ext(const struct ggml_cgraph * cgraph, const int * node_idxs, const enum ggml_op * ops, int num_ops) {
     for (int i = 0; i < num_ops; ++i) {
-        struct ggml_tensor * node = cgraph->nodes[node_idx + i];
+        if (node_idxs[i] >= cgraph->n_nodes) {
+            return false;
+        }
+
+        struct ggml_tensor * node = cgraph->nodes[node_idxs[i]];
         if (node->op != ops[i]) {
             return false;
         }
-        if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idx + i, 1)) {
+        if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idxs[i], 1)) {
             return false;
         }
         if (i > 0) {
-            struct ggml_tensor * prev = cgraph->nodes[node_idx + i - 1];
+            struct ggml_tensor * prev = cgraph->nodes[node_idxs[i - 1]];
             if (node->src[0] != prev && node->src[1] != prev) {
                 return false;
             }
@@ -525,6 +619,22 @@ static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
     return true;
 }
 
+// same as above, for sequential indices starting at node_idx
+static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, const enum ggml_op * ops, int num_ops) {
+    assert(num_ops < 32);
+
+    if (node_idx + num_ops > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+    for (int i = 0; i < num_ops; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return ggml_can_fuse_ext(cgraph, idxs, ops, num_ops);
+}
+
 #ifdef __cplusplus
 }
 #endif

data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt

@@ -5,7 +5,12 @@ find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
 message(STATUS "Metal framework found")
 
 ggml_add_backend_library(ggml-metal
-                         ggml-metal.m
+                         ggml-metal.cpp
+                         ggml-metal-device.m
+                         ggml-metal-device.cpp
+                         ggml-metal-common.cpp
+                         ggml-metal-context.m
+                         ggml-metal-ops.cpp
                         )
 
 target_link_libraries(ggml-metal PRIVATE
@@ -18,10 +23,6 @@ if (GGML_METAL_NDEBUG)
     add_compile_definitions(GGML_METAL_NDEBUG)
 endif()
 
-if (GGML_METAL_USE_BF16)
-    add_compile_definitions(GGML_METAL_USE_BF16)
-endif()
-
 # copy metal files to bin directory
 configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
 configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
@@ -71,7 +72,9 @@ else()
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-        set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        # note: adding -g causes segmentation fault during compile
+        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
@@ -90,7 +93,7 @@ else()
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-            xcrun -sdk macosx metallib - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
         DEPENDS ggml-metal.metal ${METALLIB_COMMON}