matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-17 19:27:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: update vendor code to commit ba1cb19c (#8101)

Browse Source
This commit is contained in:
Jeffrey Morgan
2024-12-14 14:55:51 -08:00
committed by GitHub
parent 60f75560a2
commit 7a81daf026
273 changed files with 3194 additions and 1900 deletions
Download Patch File Download Diff File Expand all files Collapse all files

72
llama/ggml-cuda/common.cuh vendored
View File

@@ -1,5 +1,5 @@
/**
* llama.cpp - commit 40c6d79fb52f995f47507fedfeaae2ac05d9b35c - do not edit this file
* llama.cpp - commit ba1cb19cdd0d92e012e0f6e009e0620f854b6afd - do not edit this file
*
* MIT License
*
@@ -67,28 +67,28 @@
#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
#define CUDART_HMASK 12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
#define CC_PASCAL 600
#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
#define CC_VOLTA 700
#define CC_TURING 750
#define CC_AMPERE 800
#define CC_OFFSET_AMD 1000000
#define GGML_CUDA_CC_PASCAL 600
#define GGML_CUDA_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
#define GGML_CUDA_CC_VOLTA 700
#define GGML_CUDA_CC_TURING 750
#define GGML_CUDA_CC_AMPERE 800
#define GGML_CUDA_CC_OFFSET_AMD 1000000
// GCN/CNDA, wave size is 64
#define CC_GCN4 (CC_OFFSET_AMD + 803) // Tonga, Fiji, Polaris, minimum for fast fp16
#define CC_VEGA (CC_OFFSET_AMD + 900) // Vega56/64, minimum for fp16 dual issue
#define CC_VEGA20 (CC_OFFSET_AMD + 906) // MI50/Radeon VII, minimum for dp4a
#define CC_CDNA (CC_OFFSET_AMD + 908) // MI100, minimum for MFMA, acc registers
#define CC_CDNA2 (CC_OFFSET_AMD + 910) // MI210, minimum acc register renameing
#define CC_CDNA3 (CC_OFFSET_AMD + 942) // MI300
#define GGML_CUDA_CC_GCN4 (GGML_CUDA_CC_OFFSET_AMD + 803) // Tonga, Fiji, Polaris, minimum for fast fp16
#define GGML_CUDA_CC_VEGA (GGML_CUDA_CC_OFFSET_AMD + 900) // Vega56/64, minimum for fp16 dual issue
#define GGML_CUDA_CC_VEGA20 (GGML_CUDA_CC_OFFSET_AMD + 906) // MI50/Radeon VII, minimum for dp4a
#define GGML_CUDA_CC_CDNA (GGML_CUDA_CC_OFFSET_AMD + 908) // MI100, minimum for MFMA, acc registers
#define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 910) // MI210, minimum acc register renameing
#define GGML_CUDA_CC_CDNA3 (GGML_CUDA_CC_OFFSET_AMD + 942) // MI300
// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
#define CC_RDNA1 (CC_OFFSET_AMD + 1010) // RX 5000
#define CC_RDNA2 (CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
#define CC_RDNA3 (CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
#define GGML_CUDA_CC_RDNA1 (GGML_CUDA_CC_OFFSET_AMD + 1010) // RX 5000
#define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
#define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
#define CC_QY1 210
#define CC_QY2 220
#define GGML_CUDA_CC_QY1 210
#define GGML_CUDA_CC_QY2 220
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
@@ -157,36 +157,36 @@ typedef float dfloat; // dequantize float
typedef float2 dfloat2;
#endif // GGML_CUDA_F16
#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
#define FP16_AVAILABLE
#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __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__ >= CC_VOLTA
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
#define FP16_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
#define INT8_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
#define FLASH_ATTN_AVAILABLE
#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
static constexpr bool fast_fp16_available(const int cc) {
return cc >= CC_PASCAL && cc != 610;
return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
}
static constexpr bool fp16_mma_available(const int cc) {
return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA;
}
static constexpr bool int8_mma_available(const int cc) {
return cc < CC_OFFSET_AMD && cc >= CC_TURING;
return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_TURING;
}
[[noreturn]]
@@ -213,7 +213,7 @@ static __device__ void no_device_code(
#endif // __CUDA_ARCH__
static __device__ __forceinline__ int warp_reduce_sum(int x) {
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
return __reduce_add_sync(0xffffffff, x);
#else
#pragma unroll
@@ -221,7 +221,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
x += __shfl_xor_sync(0xffffffff, x, offset, 32);
}
return x;
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
}
static __device__ __forceinline__ float warp_reduce_sum(float x) {
@@ -310,7 +310,7 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
}
static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
#pragma unroll
for (int offset = 16; offset > 0; offset >>= 1) {
x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
@@ -319,7 +319,7 @@ 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__ >= CC_PASCAL
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
}
#if CUDART_VERSION < CUDART_HMASK
@@ -359,13 +359,13 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if __CUDA_ARCH__ >= MIN_CC_DP4A
#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
return __dp4a(a, b, c);
#else // __CUDA_ARCH__ >= MIN_CC_DP4A
#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
const int8_t * a8 = (const int8_t *) &a;
const int8_t * b8 = (const int8_t *) &b;
return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
}