armorpaint/base/tools/iris.c/iris_safetensors.c

/*
 * iris_safetensors.c - Safetensors file format reader implementation
 */

#include "iris_safetensors.h"
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>

/* Minimal JSON parser for safetensors header */

static void skip_whitespace(const char **p) {
	while (**p == ' ' || **p == '\n' || **p == '\r' || **p == '\t')
		(*p)++;
}

static int parse_string(const char **p, char *out, size_t max_len) {
	skip_whitespace(p);
	if (**p != '"')
		return -1;
	(*p)++;

	size_t i = 0;
	while (**p && **p != '"' && i < max_len - 1) {
		if (**p == '\\') {
			(*p)++;
			if (**p == 'n')
				out[i++] = '\n';
			else if (**p == 't')
				out[i++] = '\t';
			else if (**p == 'r')
				out[i++] = '\r';
			else if (**p == '"')
				out[i++] = '"';
			else if (**p == '\\')
				out[i++] = '\\';
			else
				out[i++] = **p;
		}
		else {
			out[i++] = **p;
		}
		(*p)++;
	}
	out[i] = '\0';

	if (**p != '"')
		return -1;
	(*p)++;
	return 0;
}

static int64_t parse_int(const char **p) {
	skip_whitespace(p);
	int64_t val = 0;
	int     neg = 0;
	if (**p == '-') {
		neg = 1;
		(*p)++;
	}
	while (**p >= '0' && **p <= '9') {
		val = val * 10 + (**p - '0');
		(*p)++;
	}
	return neg ? -val : val;
}

static safetensor_dtype_t parse_dtype(const char *s) {
	if (strcmp(s, "F32") == 0)
		return DTYPE_F32;
	if (strcmp(s, "F16") == 0)
		return DTYPE_F16;
	if (strcmp(s, "BF16") == 0)
		return DTYPE_BF16;
	if (strcmp(s, "I32") == 0)
		return DTYPE_I32;
	if (strcmp(s, "I64") == 0)
		return DTYPE_I64;
	if (strcmp(s, "BOOL") == 0)
		return DTYPE_BOOL;
	return DTYPE_UNKNOWN;
}

/* Parse a tensor entry from JSON */
static int parse_tensor_entry(const char **p, safetensor_t *t) {
	skip_whitespace(p);
	if (**p != '{')
		return -1;
	(*p)++;

	t->dtype       = DTYPE_UNKNOWN;
	t->ndim        = 0;
	t->data_offset = 0;
	t->data_size   = 0;

	while (**p && **p != '}') {
		skip_whitespace(p);
		if (**p == ',') {
			(*p)++;
			continue;
		}

		char key[64];
		if (parse_string(p, key, sizeof(key)) != 0)
			return -1;

		skip_whitespace(p);
		if (**p != ':')
			return -1;
		(*p)++;
		skip_whitespace(p);

		if (strcmp(key, "dtype") == 0) {
			char dtype_str[32];
			if (parse_string(p, dtype_str, sizeof(dtype_str)) != 0)
				return -1;
			t->dtype = parse_dtype(dtype_str);
		}
		else if (strcmp(key, "shape") == 0) {
			if (**p != '[')
				return -1;
			(*p)++;
			t->ndim = 0;
			while (**p && **p != ']' && t->ndim < 8) {
				skip_whitespace(p);
				if (**p == ',') {
					(*p)++;
					continue;
				}
				t->shape[t->ndim++] = parse_int(p);
			}
			if (**p == ']')
				(*p)++;
		}
		else if (strcmp(key, "data_offsets") == 0) {
			if (**p != '[')
				return -1;
			(*p)++;
			skip_whitespace(p);
			size_t start = (size_t)parse_int(p);
			skip_whitespace(p);
			if (**p == ',')
				(*p)++;
			skip_whitespace(p);
			size_t end     = (size_t)parse_int(p);
			t->data_offset = start;
			t->data_size   = end - start;
			skip_whitespace(p);
			if (**p == ']')
				(*p)++;
		}
		else {
			/* Skip unknown value */
			if (**p == '"') {
				(*p)++;
				while (**p && **p != '"') {
					if (**p == '\\')
						(*p)++;
					if (**p)
						(*p)++;
				}
				if (**p == '"')
					(*p)++;
			}
			else if (**p == '[') {
				int depth = 1;
				(*p)++;
				while (**p && depth > 0) {
					if (**p == '[')
						depth++;
					else if (**p == ']')
						depth--;
					(*p)++;
				}
			}
			else if (**p == '{') {
				int depth = 1;
				(*p)++;
				while (**p && depth > 0) {
					if (**p == '{')
						depth++;
					else if (**p == '}')
						depth--;
					(*p)++;
				}
			}
			else {
				while (**p && **p != ',' && **p != '}')
					(*p)++;
			}
		}
	}

	if (**p == '}')
		(*p)++;
	return 0;
}

/* Parse the entire JSON header */
static int parse_header(safetensors_file_t *sf) {
	const char *p = sf->header_json;
	skip_whitespace(&p);

	if (*p != '{')
		return -1;
	p++;

	sf->num_tensors = 0;

	while (*p && *p != '}' && sf->num_tensors < SAFETENSORS_MAX_TENSORS) {
		skip_whitespace(&p);
		if (*p == ',') {
			p++;
			continue;
		}
		if (*p == '}')
			break;

		/* Parse tensor name */
		char name[256];
		if (parse_string(&p, name, sizeof(name)) != 0)
			return -1;

		skip_whitespace(&p);
		if (*p != ':')
			return -1;
		p++;

		/* Skip __metadata__ entry */
		if (strcmp(name, "__metadata__") == 0) {
			skip_whitespace(&p);
			if (*p == '{') {
				int depth = 1;
				p++;
				while (*p && depth > 0) {
					if (*p == '{')
						depth++;
					else if (*p == '}')
						depth--;
					p++;
				}
			}
			continue;
		}

		/* Parse tensor entry */
		safetensor_t *t = &sf->tensors[sf->num_tensors];
		snprintf(t->name, sizeof(t->name), "%s", name);

		if (parse_tensor_entry(&p, t) != 0)
			return -1;
		sf->num_tensors++;
	}

	return 0;
}

/* Open a safetensors file by memory-mapping it. Parses the JSON header to
 * build a tensor index (name -> offset/shape/dtype). Memory mapping lets the
 * OS page in tensor data on demand, avoiding upfront reads of multi-GB model
 * files -- only the weights actually used get loaded into RAM. */
safetensors_file_t *safetensors_open(const char *path) {
	int fd = open(path, O_RDONLY);
	if (fd < 0) {
		perror("safetensors_open: open failed");
		return NULL;
	}

	struct stat st;
	if (fstat(fd, &st) < 0) {
		perror("safetensors_open: fstat failed");
		close(fd);
		return NULL;
	}

	size_t file_size = (size_t)st.st_size;
	if (file_size < 8) {
		fprintf(stderr, "safetensors_open: file too small\n");
		close(fd);
		return NULL;
	}

	void *data = mmap(NULL, file_size, PROT_READ, MAP_PRIVATE, fd, 0);
	close(fd);

	if (data == MAP_FAILED) {
		perror("safetensors_open: mmap failed");
		return NULL;
	}

	/* Read header size (8-byte little-endian) */
	uint64_t header_size = 0;
	memcpy(&header_size, data, 8);

	if (header_size > file_size - 8) {
		fprintf(stderr, "safetensors_open: invalid header size\n");
		munmap(data, file_size);
		return NULL;
	}

	safetensors_file_t *sf = calloc(1, sizeof(safetensors_file_t));
	if (!sf) {
		munmap(data, file_size);
		return NULL;
	}

	sf->path        = strdup(path);
	sf->data        = data;
	sf->file_size   = file_size;
	sf->header_size = (size_t)header_size;

	/* Copy header JSON for parsing */
	sf->header_json = malloc(header_size + 1);
	if (!sf->header_json) {
		safetensors_close(sf);
		return NULL;
	}
	memcpy(sf->header_json, (char *)data + 8, header_size);
	sf->header_json[header_size] = '\0';

	/* Parse header */
	if (parse_header(sf) != 0) {
		fprintf(stderr, "safetensors_open: failed to parse header\n");
		safetensors_close(sf);
		return NULL;
	}

	/* Validate that all tensor data fits within the file.
	 * Catches truncated downloads before they cause segfaults. */
	size_t data_region = file_size - 8 - (size_t)header_size;
	for (int i = 0; i < sf->num_tensors; i++) {
		safetensor_t *t = &sf->tensors[i];
		if (t->data_offset + t->data_size > data_region) {
			fprintf(stderr,
			        "safetensors_open: %s: truncated file\n"
			        "  tensor '%s' needs data at offset %zu-%zu, "
			        "but file only has %zu bytes of tensor data.\n"
			        "  The file may be from an interrupted download. "
			        "Re-download it.\n",
			        path, t->name, t->data_offset, t->data_offset + t->data_size, data_region);
			safetensors_close(sf);
			return NULL;
		}
	}

	return sf;
}

void safetensors_close(safetensors_file_t *sf) {
	if (!sf)
		return;
	if (sf->data)
		munmap(sf->data, sf->file_size);
	free(sf->path);
	free(sf->header_json);
	free(sf);
}

const safetensor_t *safetensors_find(const safetensors_file_t *sf, const char *name) {
	for (int i = 0; i < sf->num_tensors; i++) {
		if (strcmp(sf->tensors[i].name, name) == 0) {
			return &sf->tensors[i];
		}
	}
	return NULL;
}

const void *safetensors_data(const safetensors_file_t *sf, const safetensor_t *t) {
	size_t offset = 8 + sf->header_size + t->data_offset;
	return (const char *)sf->data + offset;
}

int64_t safetensor_numel(const safetensor_t *t) {
	int64_t n = 1;
	for (int i = 0; i < t->ndim; i++) {
		n *= t->shape[i];
	}
	return n;
}

/* Convert BF16 to F32 */
static float bf16_to_f32(uint16_t bf16) {
	uint32_t f32 = ((uint32_t)bf16) << 16;
	float    result;
	memcpy(&result, &f32, sizeof(float));
	return result;
}

/* Convert F16 to F32 */
static float f16_to_f32(uint16_t f16) {
	uint32_t sign = (f16 >> 15) & 0x1;
	uint32_t exp  = (f16 >> 10) & 0x1F;
	uint32_t mant = f16 & 0x3FF;

	uint32_t f32;
	if (exp == 0) {
		if (mant == 0) {
			f32 = sign << 31;
		}
		else {
			/* Denormalized number */
			exp = 1;
			while ((mant & 0x400) == 0) {
				mant <<= 1;
				exp--;
			}
			mant &= 0x3FF;
			f32 = (sign << 31) | ((exp + 127 - 15) << 23) | (mant << 13);
		}
	}
	else if (exp == 31) {
		/* Inf or NaN */
		f32 = (sign << 31) | 0x7F800000 | (mant << 13);
	}
	else {
		f32 = (sign << 31) | ((exp + 127 - 15) << 23) | (mant << 13);
	}

	float result;
	memcpy(&result, &f32, sizeof(float));
	return result;
}

/* Return a tensor's data as a newly allocated f32 array. Handles dtype
 * conversion: f32 is memcpy'd directly, bf16 and f16 are converted
 * element-wise. The caller owns the returned buffer. This is the main
 * entry point for loading individual weights during model initialization. */
float *safetensors_get_f32(const safetensors_file_t *sf, const safetensor_t *t) {
	int64_t n   = safetensor_numel(t);
	float  *out = malloc(n * sizeof(float));
	if (!out)
		return NULL;

	const void *data = safetensors_data(sf, t);

	switch (t->dtype) {
	case DTYPE_F32:
		memcpy(out, data, n * sizeof(float));
		break;

	case DTYPE_F16: {
		const uint16_t *src = (const uint16_t *)data;
		for (int64_t i = 0; i < n; i++) {
			out[i] = f16_to_f32(src[i]);
		}
		break;
	}

	case DTYPE_BF16: {
		const uint16_t *src = (const uint16_t *)data;
		for (int64_t i = 0; i < n; i++) {
			out[i] = bf16_to_f32(src[i]);
		}
		break;
	}

	case DTYPE_Q8_0: {
		/* GGML Q8_0: blocks of 32 elements, each block is an fp16 scale
		 * (2 bytes) followed by 32 signed int8 quants -> 34 bytes. */
		const uint8_t *src = (const uint8_t *)data;
		int64_t        nb  = n / 32;
		for (int64_t b = 0; b < nb; b++) {
			const uint8_t *blk = src + b * 34;
			uint16_t       scale_bits;
			memcpy(&scale_bits, blk, sizeof(scale_bits));
			float         scale = f16_to_f32(scale_bits);
			const int8_t *qs    = (const int8_t *)(blk + 2);
			for (int j = 0; j < 32; j++) {
				out[b * 32 + j] = (float)qs[j] * scale;
			}
		}
		break;
	}

	default:
		fprintf(stderr, "safetensors_get_f32: unsupported dtype\n");
		free(out);
		return NULL;
	}

	return out;
}

int safetensor_is_bf16(const safetensor_t *t) {
	return t && t->dtype == DTYPE_BF16;
}

int safetensor_is_q8_0(const safetensor_t *t) {
	return t && t->dtype == DTYPE_Q8_0;
}

uint16_t *safetensors_get_bf16(const safetensors_file_t *sf, const safetensor_t *t) {
	if (!sf || !t)
		return NULL;

	/* Only works for BF16 tensors */
	if (t->dtype != DTYPE_BF16) {
		fprintf(stderr, "safetensors_get_bf16: tensor is not BF16 (dtype=%d)\n", t->dtype);
		return NULL;
	}

	int64_t n = safetensor_numel(t);
	if (n <= 0)
		return NULL;

	const void *data = safetensors_data(sf, t);
	if (!data)
		return NULL;

	/* Allocate and copy bf16 data (already in correct format) */
	uint16_t *out = (uint16_t *)malloc(n * sizeof(uint16_t));
	if (!out)
		return NULL;

	memcpy(out, data, n * sizeof(uint16_t));
	return out;
}

/* Get direct pointer to bf16 data in mmap'd region (no copy, caller must not free) */
uint16_t *safetensors_get_bf16_direct(const safetensors_file_t *sf, const safetensor_t *t) {
	if (!sf || !t)
		return NULL;
	if (t->dtype != DTYPE_BF16)
		return NULL;
	return (uint16_t *)safetensors_data(sf, t);
}

void safetensor_print(const safetensor_t *t) {
	const char *dtype_names[] = {"F32", "F16", "BF16", "I32", "I64", "BOOL", "Q8_0"};
	const char *dtype_name    = t->dtype >= 0 && t->dtype <= 6 ? dtype_names[t->dtype] : "UNKNOWN";

	printf("%s: dtype=%s, shape=[", t->name, dtype_name);
	for (int i = 0; i < t->ndim; i++) {
		printf("%ld%s", (long)t->shape[i], i < t->ndim - 1 ? ", " : "");
	}
	printf("], offset=%zu, size=%zu\n", t->data_offset, t->data_size);
}

void safetensors_print_all(const safetensors_file_t *sf) {
	printf("Safetensors file: %s\n", sf->path);
	printf("File size: %zu bytes\n", sf->file_size);
	printf("Header size: %zu bytes\n", sf->header_size);
	printf("Number of tensors: %d\n\n", sf->num_tensors);

	for (int i = 0; i < sf->num_tensors; i++) {
		safetensor_print(&sf->tensors[i]);
	}
}