From ea1aeba48b42f5f6ec41412ef271f5d708b0fa2f Mon Sep 17 00:00:00 2001 From: S Date: Mon, 8 Apr 2024 22:47:46 +0100 Subject: [PATCH] dranger003: Fix more int overflow during quant. --- ggml-cuda.cu | 6 ++-- ggml-cuda/convert.cu | 74 +++++++++++++++++++++--------------------- ggml-cuda/convert.cuh | 2 +- ggml-cuda/dmmv.cu | 2 +- ggml-cuda/quantize.cu | 16 ++++----- ggml-cuda/quantize.cuh | 2 +- 6 files changed, 51 insertions(+), 51 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index f51b2042d..5616e2348 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -1225,7 +1225,7 @@ static void ggml_cuda_op_mul_mat_cublas( // the main device has a larger memory buffer to hold the results from all GPUs // ldc == nrows of the matrix that cuBLAS writes into - int ldc = id == ctx.device ? ne0 : row_diff; + int64_t ldc = id == ctx.device ? ne0 : row_diff; const int compute_capability = ggml_cuda_info().devices[id].cc; @@ -1377,8 +1377,8 @@ static void ggml_cuda_op_mul_mat( const int64_t ne0 = dst->ne[0]; const int64_t ne1 = dst->ne[1]; - const int nb2 = dst->nb[2]; - const int nb3 = dst->nb[3]; + const int64_t nb2 = dst->nb[2]; + const int64_t nb3 = dst->nb[3]; GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer)); GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer)); diff --git a/ggml-cuda/convert.cu b/ggml-cuda/convert.cu index 18a31edc3..ed4fa2748 100644 --- a/ggml-cuda/convert.cu +++ b/ggml-cuda/convert.cu @@ -4,14 +4,14 @@ #define CUDA_Q8_0_NE_ALIGN 2048 template -static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) { - const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x); +static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) { + const int64_t i = 2*(blockDim.x*blockIdx.x + threadIdx.x); if (i >= k) { return; } - const int ib = i/qk; // block index + const int64_t ib = i/qk; // block index const int iqs = (i%qk)/qr; // quant index const int iybs = i - i%qk; // y block start index const int y_offset = qr == 1 ? 1 : qk/2; @@ -25,7 +25,7 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __ } template -static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int k) { +static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) { #if __CUDA_ARCH__ >= CC_PASCAL constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE; @@ -68,13 +68,13 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h template static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; // assume 32 threads const int tid = threadIdx.x; const int il = tid/8; const int ir = tid%8; - const int ib = 8*i + ir; + const int64_t ib = 8*i + ir; if (ib >= nb32) { return; } @@ -96,13 +96,13 @@ static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t template static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; // assume 32 threads const int tid = threadIdx.x; const int il = tid/8; const int ir = tid%8; - const int ib = 8*i + ir; + const int64_t ib = 8*i + ir; if (ib >= nb32) { return; } @@ -313,14 +313,14 @@ template static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { const block_q6_K * x = (const block_q6_K *) vx; - const int i = blockIdx.x; + const int64_t i = blockIdx.x; #if QK_K == 256 // assume 64 threads - this is very slightly better than the one below - const int tid = threadIdx.x; - const int ip = tid/32; // ip is 0 or 1 - const int il = tid - 32*ip; // 0...32 - const int is = 8*ip + il/16; + const int64_t tid = threadIdx.x; + const int64_t ip = tid/32; // ip is 0 or 1 + const int64_t il = tid - 32*ip; // 0...32 + const int64_t is = 8*ip + il/16; dst_t * y = yy + i*QK_K + 128*ip + il; @@ -337,9 +337,9 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t #else // assume 32 threads - const int tid = threadIdx.x; - const int ip = tid/16; // 0 or 1 - const int il = tid - 16*ip; // 0...15 + const int64_t tid = threadIdx.x; + const int64_t ip = tid/16; // 0 or 1 + const int64_t il = tid - 16*ip; // 0...15 dst_t * y = yy + i*QK_K + 16*ip + il; @@ -571,12 +571,12 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst #endif template -static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { +static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) { const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE); dequantize_block<<>>(vx, y, k); } -static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int k, cudaStream_t stream) { +static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) { const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN; if (k % CUDA_Q8_0_NE_ALIGN == 0) { const bool need_check = false; @@ -588,7 +588,7 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * } template -static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; #if QK_K == 256 dequantize_block_q2_K<<>>(vx, y); @@ -598,7 +598,7 @@ static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cu } template -static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; #if QK_K == 256 dequantize_block_q3_K<<>>(vx, y); @@ -608,27 +608,27 @@ static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cu } template -static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb32 = k / 32; const int nb = (k + 255) / 256; dequantize_block_q4_0<<>>(vx, y, nb32); } template -static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb32 = k / 32; const int nb = (k + 255) / 256; dequantize_block_q4_1<<>>(vx, y, nb32); } template -static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_q4_K<<>>(vx, y); } template -static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; #if QK_K == 256 dequantize_block_q5_K<<>>(vx, y); @@ -638,7 +638,7 @@ static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cu } template -static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; #if QK_K == 256 dequantize_block_q6_K<<>>(vx, y); @@ -648,55 +648,55 @@ static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cu } template -static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq2_xxs<<>>(vx, y); } template -static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq2_xs<<>>(vx, y); } template -static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq2_s<<>>(vx, y); } template -static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq3_xxs<<>>(vx, y); } template -static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq3_s<<>>(vx, y); } template -static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq1_s<<>>(vx, y); } template -static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = (k + QK_K - 1) / QK_K; dequantize_block_iq4_nl<<>>(vx, y); } template -static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; dequantize_block_iq1_m<<>>(vx, y); } template -static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { +static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = (k + QK_K - 1) / QK_K; #if QK_K == 64 dequantize_block_iq4_nl<<>>(vx, y); @@ -706,8 +706,8 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, } template -static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; +static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) { + const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x; if (i >= k) { return; @@ -719,7 +719,7 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res } template -static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { +static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) { const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE; convert_unary<<>>(vx, y, k); } diff --git a/ggml-cuda/convert.cuh b/ggml-cuda/convert.cuh index db34c0be9..5394be9f1 100644 --- a/ggml-cuda/convert.cuh +++ b/ggml-cuda/convert.cuh @@ -3,7 +3,7 @@ #define CUDA_DEQUANTIZE_BLOCK_SIZE 256 template -using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream); +using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream); typedef to_t_cuda_t to_fp32_cuda_t; typedef to_t_cuda_t to_fp16_cuda_t; diff --git a/ggml-cuda/dmmv.cu b/ggml-cuda/dmmv.cu index 3097a9510..7313e3e17 100644 --- a/ggml-cuda/dmmv.cu +++ b/ggml-cuda/dmmv.cu @@ -577,7 +577,7 @@ template static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) { // qk = quantized weights per x block // qr = number of quantized weights per data value in x block - const int row = blockIdx.x*blockDim.y + threadIdx.y; + const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y; if (row >= nrows) { return; diff --git a/ggml-cuda/quantize.cu b/ggml-cuda/quantize.cu index a1fbc9932..7578c4b6c 100644 --- a/ggml-cuda/quantize.cu +++ b/ggml-cuda/quantize.cu @@ -1,20 +1,20 @@ #include "quantize.cuh" -static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) { - const int ix = blockDim.x*blockIdx.x + threadIdx.x; +static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx_padded) { + const int64_t ix = (int64_t)blockDim.x*blockIdx.x + threadIdx.x; if (ix >= kx_padded) { return; } - const int iy = blockDim.y*blockIdx.y + threadIdx.y; + const int64_t iy = (int64_t)blockDim.y*blockIdx.y + threadIdx.y; - const int i_padded = iy*kx_padded + ix; + const int64_t i_padded = (int64_t)iy*kx_padded + ix; block_q8_1 * y = (block_q8_1 *) vy; - const int ib = i_padded / QK8_1; // block index - const int iqs = i_padded % QK8_1; // quant index + const int64_t ib = i_padded / QK8_1; // block index + const int64_t iqs = i_padded % QK8_1; // quant index const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; float amax = fabsf(xi); @@ -36,8 +36,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest reinterpret_cast(y[ib].ds.y) = sum; } -void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) { - const int block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE; +void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream) { + const int64_t block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE; const dim3 num_blocks(block_num_x, ky, 1); const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1); quantize_q8_1<<>>(x, vy, kx, kx_padded); diff --git a/ggml-cuda/quantize.cuh b/ggml-cuda/quantize.cuh index adb89c83a..b37a4752f 100644 --- a/ggml-cuda/quantize.cuh +++ b/ggml-cuda/quantize.cuh @@ -2,4 +2,4 @@ #define CUDA_QUANTIZE_BLOCK_SIZE 256 -void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream); +void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream);