Add more choices to quantization tool. Post processing after sim_anneal(). (optimizer.py/ext_quant.cpp)

exllamav2/exllamav2_ext/cpp/sampling.cpp

@@ -38,7 +38,7 @@ void apply_rep_penalty_cpu
 //    {
 //        if (g_rep_mask) free(g_rep_mask);
 //        g_vocab_size = vocab_size;
-//        g_rep_mask = (bool*) malloc(g_vocab_size * sizeof(bool));
+//        g_rep_mask = (bool*) calloc(1, g_vocab_size * sizeof(bool));
 //    }
 //    memset(g_rep_mask, 0, g_vocab_size * sizeof(bool));
     bool* g_rep_mask = (bool*) calloc(vocab_size, sizeof(bool));
@@ -655,7 +655,7 @@ int tfs_cpu
 
     int nc = sort_descending(num_candidates, temp_probs, temp_indices, num_candidates);
 
-    float* derivative = (float*) malloc(nc * sizeof(float));
+    float* derivative = (float*) calloc(1, nc * sizeof(float));
     float dsum = 0.0f;
     for (int i = 0; i < nc - 2; i++)
     {
@@ -759,9 +759,9 @@ int typical_cpu
 
     int r_candidates = pre_sort_descending(num_candidates, temp_probs, temp_indices);
 
-    float* temp = (float*) malloc(r_candidates * sizeof(float));
-    int* entropy_dev_order = (int*) malloc(r_candidates * sizeof(int));
-    int* temp_indices_2 = (int*) malloc(r_candidates * sizeof(int));
+    float* temp = (float*) calloc(1, r_candidates * sizeof(float));
+    int* entropy_dev_order = (int*) calloc(1, r_candidates * sizeof(int));
+    int* temp_indices_2 = (int*) calloc(1, r_candidates * sizeof(int));
 
     float neg_entropy = 0.0f;
     for (int i = 0; i < r_candidates; i++)

exllamav2/exllamav2_ext/cuda/cache.cu

@@ -168,8 +168,8 @@ __global__ void fp16_to_q_kv_paged_kernel
 
     int page = block_table[pages_per_seq * y + x];
     int seqlen = cache_seqlens[y];
-    int vx_a = page_size * x;
-    int px_a = seqlen - vx_a;
+    int vx_a = (int64_t)page_size * x;
+    int px_a = (int64_t)seqlen - vx_a;
     int px_b = px_a + q_len;
 
     if (dim % BLOCKSIZE_Q)
@@ -349,7 +349,7 @@ __global__ void q_to_fp16_kv_paged_kernel
     int seqlen = cache_seqlens[y];
     if (!seqlen) return;
 
-    int vx_a = page_size * x;
+    int vx_a = (int64_t)page_size * x;
     int vx_b = min(vx_a + page_size, seqlen);
 
     if (dim < BLOCKSIZE_Q)

exllamav2/exllamav2_ext/cuda/q_matrix.cu

@@ -603,9 +603,18 @@ bool QMatrix::make_sequential(const uint32_t* cpu_g_idx, cudaStream_t stream)
         return false;
     }
 
+    // Zero out the allocated memory
+    size_t mem_size = (height / 8) * width * sizeof(uint32_t);
+    err = cudaMemset(cuda_new_qweight, 0, mem_size);
+    if (err != cudaSuccess) {;;;
+        printf("CUDA memset failed: %s\n", cudaGetErrorString(err));
+        cudaFree(cuda_new_qweight);  // Free the allocated memory in case of error
+        return err;
+    }
+
     uint32_t* cpu_g_idx_map = (uint32_t*) calloc(groups, sizeof(uint32_t));
-    uint32_t* cpu_x_map = (uint32_t*) malloc(height * sizeof(uint32_t));
-    uint32_t* cpu_x_map_inv = (uint32_t*) malloc(height * sizeof(uint32_t));
+    uint32_t* cpu_x_map = (uint32_t*) calloc(1, height * sizeof(uint32_t));
+    uint32_t* cpu_x_map_inv = (uint32_t*) calloc(1, height * sizeof(uint32_t));
 
     // Group histogram
 

exllamav2/exllamav2_ext/cuda/util.cu

@@ -2,7 +2,7 @@
 
 void print_global_mem(const half* ptr, int rows, int columns, int stride)
 {
-    half* temp = (half*) malloc(rows * columns * sizeof(half));
+    half* temp = (half*) calloc(1, rows * columns * sizeof(half));
 
     cudaDeviceSynchronize();
     cudaMemcpyAsync(temp, ptr, rows * columns * sizeof(half), cudaMemcpyDeviceToHost);