whispercpp 1.3.2 → 1.3.4

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

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-cuda.h"
 
 #include <cstdint>
@@ -19,10 +20,10 @@
 #endif
 #include "ggml-common.h"
 
-#include <cstdio>
 #include <array>
 #include <cassert>
 #include <cfloat>
+#include <cstdio>
 #include <string>
 #include <vector>
 
@@ -56,7 +57,7 @@
 #define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
 #define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
 #define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
-#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
+#define GGML_CUDA_CC_CDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x910)  // MI210, minimum acc register renameing
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
 
@@ -72,21 +73,28 @@
 #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
 #define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
-#define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
-#define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
+#define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_CDNA1(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
+#define GGML_CUDA_CC_IS_CDNA2(cc) (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
+#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
-#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
+#define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
 
-#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
 #define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#    define GGML_CUDA_USE_CUB
+#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+
 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
     return false;
@@ -101,9 +109,9 @@ constexpr bool ggml_cuda_has_arch(const int arch) {
     return ggml_cuda_has_arch_impl(arch, __CUDA_ARCH_LIST__);
 }
 
-constexpr int ggml_cuda_highest_compiled_arch_impl(const int arch, const int cur) {
+constexpr int ggml_cuda_highest_compiled_arch_impl(const int /*arch*/, const int cur) {
     if (cur == 0) {
-        GGML_ABORT("ggml was not compiled with any CUDA arch <= %d", arch);
+        return -1;
     }
     return cur;
 }
@@ -177,51 +185,68 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes)                                                       \
+        do {                                                                                                   \
+            static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = { false };                         \
+            const int   id                                                = ggml_cuda_get_device();            \
+            if (!shared_memory_limit_raised[id]) {                                                             \
+                CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
+                shared_memory_limit_raised[id] = true;                                                         \
+            }                                                                                                  \
+        } while (0)
+#else
+#    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
+        do {                                             \
+            GGML_UNUSED(nbytes);                         \
+        } while (0)
+#endif // !(defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else
 #define GGML_CUDA_ASSUME(x)
 #endif // CUDART_VERSION >= 11010
 
-#ifdef GGML_CUDA_F16
-typedef half dfloat; // dequantize float
-typedef half2 dfloat2;
-#else
-typedef float dfloat; // dequantize float
-typedef float2 dfloat2;
-#endif // GGML_CUDA_F16
-
 #if (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 
-#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
+
+#if defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
+#define AMD_MFMA_AVAILABLE
+#endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
+
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#define TURING_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
-#define NEW_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#define AMPERE_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #define CP_ASYNC_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
-#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 #define FLASH_ATTN_AVAILABLE
-#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
@@ -233,54 +258,101 @@ static bool fast_fp16_available(const int cc) {
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fast_fp16_hardware_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }
 
 // Any FP16 tensor core instructions are available for ggml code.
 static bool fp16_mma_available(const int cc) {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+#if defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
     return false;
 #else
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+    if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) ||
+        GGML_CUDA_CC_IS_MTHREADS(cc)) {
+        return true;
+    } else if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+#if defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+        return true;
+#else
+        return false;
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+    } else {
+        return false;
+    }
+#endif // defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
     return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+}
+
+static bool bf16_mma_hardware_available(const int cc) {
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+}
+
+static bool fp32_mma_hardware_available(const int cc) {
+    return GGML_CUDA_CC_IS_CDNA(cc);
+}
+
+static bool amd_mfma_available(const int cc) {
+#if !defined(GGML_HIP_NO_MMQ_MFMA)
+    return GGML_CUDA_CC_IS_CDNA(cc);
+#else
+    return false;
+#endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
-static bool new_mma_available(const int cc) {
+static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }
 
+static bool ampere_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+}
+
 static bool cp_async_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
+// Maximum number of bytes that can be copied in a single instruction.
+static constexpr __device__ int ggml_cuda_get_max_cpy_bytes() {
+#ifdef GGML_USE_HIP
+    return 16;
+#else
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+    return 16;
+#else
+    return 8;
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // GGML_USE_HIP
+}
+
+
 [[noreturn]]
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
     printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
            file_name, line, function_name, arch);
     GGML_UNUSED(arch_list);
 #else
     printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
            file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
@@ -317,7 +389,7 @@ struct ggml_cuda_unroll<1> {
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
@@ -325,7 +397,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
         x += __shfl_xor_sync(0xffffffff, x, offset, width);
     }
     return x;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 }
 
 template<int width = WARP_SIZE>
@@ -362,6 +434,32 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ int warp_reduce_all(int x) {
+    if (width == ggml_cuda_get_physical_warp_size()) {
+        return __all_sync(0xffffffff, x);
+    } else {
+#pragma unroll
+        for (int offset = width/2; offset > 0; offset >>= 1) {
+            x = __shfl_xor_sync(0xffffffff, x, offset, width) && x;
+        }
+        return x;
+    }
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ int warp_reduce_any(int x) {
+    if (width == ggml_cuda_get_physical_warp_size()) {
+        return __any_sync(0xffffffff, x);
+    } else {
+#pragma unroll
+        for (int offset = width/2; offset > 0; offset >>= 1) {
+            x = __shfl_xor_sync(0xffffffff, x, offset, width) || x;
+        }
+        return x;
+    }
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -374,11 +472,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !defined(GGML_USE_HIP) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !defined(GGML_USE_HIP) && CUDART_VERSION < CUDART_HMAX
 
 #else
    NO_DEVICE_CODE;
@@ -388,25 +486,21 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if defined(GGML_USE_HIP) && HIP_VERSION >= 50700000
+#if defined(GGML_USE_HIP)
     return half2(__hmax(a.x, b.x), __hmax(a.y, b.y));
-#elif !defined(GGML_USE_HIP) && CUDART_VERSION >= CUDART_HMAX
+#elif CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
-#elif !defined(GGML_USE_HIP)
+#else
     half2 ret;
     reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
     reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
     return ret;
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
 #endif
 }
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 #pragma unroll
    for (int offset = width/2; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, width));
@@ -415,19 +509,20 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 }
 
-#if CUDART_VERSION < CUDART_HMASK
+#if (defined(CUDART_VERSION) && CUDART_VERSION < CUDART_HMASK) || defined(GGML_USE_HIP) || \
+    (defined(MUSART_VERSION) && MUSART_VERSION < MUSART_HMASK)
 static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
     const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
     const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
     return mask_low | mask_high;
 }
-#endif // CUDART_VERSION < CUDART_HMASK
+#endif // (defined(CUDART_VERSION) && CUDART_VERSION < CUDART_HMASK) || defined(GGML_USE_HIP) || (defined(MUSART_VERSION) && MUSART_VERSION < MUSART_HMASK)
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3) || defined(RDNA4)
@@ -453,7 +548,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif
     return c;
 
-#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP)
 
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
     return __dp4a(a, b, c);
@@ -463,13 +558,134 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
 
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
+}
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float v, const float u) {
+    acc += v*u;
+}
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v, const float2 u) {
+    acc += v.x*u.x;
+    acc += v.y*u.y;
+}
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
+#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+    asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
+#else
+#ifdef FAST_FP16_AVAILABLE
+    const float2 tmp = __half22float2(v*u);
+    acc += tmp.x + tmp.y;
+#else
+    const float2 tmpv = __half22float2(v);
+    const float2 tmpu = __half22float2(u);
+    acc += tmpv.x * tmpu.x;
+    acc += tmpv.y * tmpu.y;
+#endif // FAST_FP16_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
+}
+
+static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {
+#ifdef FAST_FP16_AVAILABLE
+    acc += v*u;
+#else
+    const float2 tmpv = __half22float2(v);
+    const float2 tmpu = __half22float2(u);
+    float2 tmpacc = __half22float2(acc);
+    tmpacc.x += tmpv.x * tmpu.x;
+    tmpacc.y += tmpv.y * tmpu.y;
+    acc = make_half2(tmpacc.x, tmpacc.y);
+#endif // FAST_FP16_AVAILABLE
+}
+
+// Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
+template <int nbytes, int alignment = 0>
+static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
+    if constexpr (alignment != 0) {
+        static_assert(nbytes % alignment == 0, "bad alignment");
+    }
+    constexpr int nb_per_cpy = alignment == 0 ? nbytes : alignment;
+
+#pragma unroll
+    for (int i = 0; i < nbytes/nb_per_cpy; ++i) {
+        if constexpr (nb_per_cpy == 1) {
+            ((char *) dst)[i] = ((const char *) src)[i];
+        } else if constexpr (nb_per_cpy == 2) {
+            ((short *) dst)[i] = ((const short *) src)[i];
+        } else if constexpr (nb_per_cpy == 4) {
+            ((int *) dst)[i] = ((const int *) src)[i];
+        } else if constexpr (nb_per_cpy == 8) {
+            ((int2 *) dst)[i] = ((const int2 *) src)[i];
+        } else if constexpr (nb_per_cpy == 16) {
+            ((int4 *) dst)[i] = ((const int4 *) src)[i];
+        } else {
+            static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
+        }
+    }
+}
+
+static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
+#if CUDART_VERSION >= 12080
+    const nv_bfloat16 e = __nv_cvt_e8m0_to_bf16raw(x);
+    return (float) e;
+#else
+    uint32_t bits;
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    float result;
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+#endif // CUDART_VERSION >= 12050
+}
+
+// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
+// Precompute mp (m' in the paper) and L such that division
+// can be computed using a multiply (high 32b of 64b result)
+// and a shift:
+//
+// n/d = (mulhi(n, mp) + n) >> L;
+static const uint3 init_fastdiv_values(uint32_t d) {
+    GGML_ASSERT(d != 0);
+
+    // compute L = ceil(log2(d));
+    uint32_t L = 0;
+    while (L < 32 && (uint32_t{ 1 } << L) < d) {
+        L++;
+    }
+
+    uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);
+    // pack divisor as well to reduce error surface
+    return make_uint3(mp, L, d);
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+static __device__ __forceinline__ uint32_t fastdiv(uint32_t n, const uint3 fastdiv_values) {
+    // expects fastdiv_values to contain <mp, L, divisor> in <x, y, z>
+    // fastdiv_values.z is unused and optimized away by the compiler.
+    // Compute high 32 bits of n * mp
+    const uint32_t hi = __umulhi(n, fastdiv_values.x);
+    // add n, apply bit shift
+    return (hi + n) >> fastdiv_values.y;
+}
+
+static __device__ __forceinline__ uint32_t fastmodulo(uint32_t n, const uint3 fastdiv_values) {
+    // expects  fastdiv_values to contain <mp, L, divisor> in <x, y, z> (see init_fastdiv_values)
+    return n - fastdiv(n, fastdiv_values) * fastdiv_values.z;
+}
 
-typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+// Calculate both division and modulo at once, returns <n/divisor, n%divisor>
+static __device__ __forceinline__ uint2 fast_div_modulo(uint32_t n, const uint3 fastdiv_values) {
+    // expects  fastdiv_values to contain <mp, L, divisor> in <x, y, z> (see init_fastdiv_values)
+    const uint32_t div_val = fastdiv(n, fastdiv_values);
+    const uint32_t mod_val = n - div_val * fastdiv_values.z;
+    return make_uint2(div_val, mod_val);
+}
+
+typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, float2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
     const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
@@ -527,6 +743,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
     static constexpr int qi = QI8_0;
 };
 
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_MXFP4> {
+    static constexpr int qk = QK_MXFP4;
+    static constexpr int qr = QR_MXFP4;
+    static constexpr int qi = QI_MXFP4;
+};
+
 template<>
 struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
     static constexpr int qk = QK_K;
@@ -635,6 +858,7 @@ struct ggml_cuda_device_info {
         int     nsm;                // number of streaming multiprocessors
         size_t  smpb;               // max. shared memory per block
         size_t  smpbo;              // max. shared memory per block (with opt-in)
+        bool    integrated;         // Device is integrated as opposed to discrete
         bool    vmm;                // virtual memory support
         size_t  vmm_granularity;    // granularity of virtual memory
         size_t  total_vram;
@@ -711,7 +935,7 @@ struct ggml_tensor_extra_gpu {
 };
 
 
-#if (defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS))
+#if (defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)) || defined(GGML_MUSA_GRAPHS)
 #define USE_CUDA_GRAPH
 #endif
 
@@ -769,21 +993,7 @@ struct ggml_backend_cuda_context {
         name(GGML_CUDA_NAME + std::to_string(device)) {
     }
 
-    ~ggml_backend_cuda_context() {
-        if (copy_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(copy_event));
-        }
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
-            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
-                if (streams[i][j] != nullptr) {
-                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
-                }
-            }
-            if (cublas_handles[i] != nullptr) {
-                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
-            }
-        }
-    }
+    ~ggml_backend_cuda_context();
 
     cudaStream_t stream(int device, int stream) {
         if (streams[device][stream] == nullptr) {

data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu

@@ -34,10 +34,7 @@ static  __global__ void conv_transpose_1d_kernel(
         }
     }
     dst[global_index] = accumulator;
-    GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3);
-    GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3);
-    GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1);
-    GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2);
+    GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2);
 }
 
 static void conv_transpose_1d_f32_f32_cuda(