llama_cpp 0.2.1 → 0.3.0

data/ext/llama_cpp/src/k_quants.h

@@ -7,7 +7,13 @@
 #include <stddef.h>
 
 // Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
 #define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
 
 //
 // Super-block quantization structures
@@ -29,38 +35,67 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "w
 // weight is represented as x = a * q
 // 16 blocks of 16 elemenets each
 // Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
 typedef struct {
     uint8_t hmask[QK_K/8];     // quants - high bit
     uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
+    uint8_t scales[2];
     ggml_fp16_t d;             // super-block scale
 } block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + 11 * QK_K / 64, "wrong q3_K block size/padding");
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
 
 // 4-bit quantization
 // 16 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
 typedef struct {
     ggml_fp16_t d;             // super-block scale for quantized scales
     ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qs[QK_K/2];        // 4--bit quants
 } block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
 
 // 5-bit quantization
 // 16 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
 typedef struct {
     ggml_fp16_t d;               // super-block scale for quantized scales
     ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
     uint8_t qh[QK_K/8];          // quants, high bit
     uint8_t qs[QK_K/2];          // quants, low 4 bits
 } block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
 
 // 6-bit quantization
 // weight is represented as x = a * q

data/ext/llama_cpp/src/llama-util.h

@@ -172,12 +172,14 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
+        // prefetch/readahead impairs performance on NUMA systems
+        if (numa) { prefetch = 0; }
 #ifdef __linux__
-        flags |= MAP_POPULATE;
+        if (prefetch) { flags |= MAP_POPULATE; }
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
@@ -191,6 +193,14 @@ struct llama_mmap {
                         strerror(errno));
             }
         }
+        if (numa) {
+            // advise the kernel not to use readahead
+            // (because the next page might not belong on the same node)
+            if (madvise(addr, file->size, MADV_RANDOM)) {
+                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
     }
 
     ~llama_mmap() {
@@ -199,7 +209,9 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true) {
+    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
+        (void) numa;
+
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
@@ -244,8 +256,10 @@ struct llama_mmap {
 #else
     static constexpr bool SUPPORTED = false;
 
-    llama_mmap(struct llama_file *, bool prefetch = true) {
-        (void)prefetch;
+    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
+        (void) prefetch;
+        (void) numa;
+
         throw std::runtime_error(std::string("mmap not supported"));
     }
 #endif