aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-15 11:56:41 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-15 11:56:41 +0200
commitb5ada41fedf1ebdaff39ee0fbef76a468b26e5f7 (patch)
tree45669fae5d4449512f79667360c5abef6ef34e42
parent588feec258dd309e4b947c25bfacad7ad956547f (diff)
downloadffwd-b5ada41fedf1ebdaff39ee0fbef76a468b26e5f7.tar.gz
ffwd-b5ada41fedf1ebdaff39ee0fbef76a468b26e5f7.zip
api: share config parsing across backends
-rw-r--r--Makefile22
-rw-r--r--embed.c148
-rw-r--r--embed_config.c141
-rw-r--r--embed_config.h8
-rw-r--r--embed_mlx.c127
5 files changed, 166 insertions, 280 deletions
diff --git a/Makefile b/Makefile
index 9843300..af88f55 100644
--- a/Makefile
+++ b/Makefile
@@ -74,6 +74,7 @@ LDFLAGS = -lm -lpthread $(BACKEND_LDFLAGS)
# Source files
SRCS = embed.c \
+ embed_config.c \
qwen_kernels.c \
qwen_kernels_generic.c \
qwen_kernels_neon.c \
@@ -164,19 +165,19 @@ TEST_CC_FLAGS = -Wall -Wextra -O2 -I.
TEST_KERNELS_SRCS = tests/test_kernels.c $(KERNEL_SRCS)
TEST_TOKENIZER_SRCS = tests/test_tokenizer.c qwen_tokenizer.c $(KERNEL_SRCS)
TEST_SAFETENSORS_SRCS = tests/test_safetensors.c qwen_safetensors.c
-TEST_BF16_SRCS = tests/test_bf16_model.c embed.c qwen_safetensors.c \
+TEST_BF16_SRCS = tests/test_bf16_model.c embed.c embed_config.c qwen_safetensors.c \
$(KERNEL_SRCS)
-TEST_QWEN3_SRCS = tests/test_qwen3.c embed.c qwen_safetensors.c \
+TEST_QWEN3_SRCS = tests/test_qwen3.c embed.c embed_config.c qwen_safetensors.c \
$(KERNEL_SRCS)
-TEST_WORKSPACE_SRCS = tests/test_workspace.c embed.c qwen_tokenizer.c \
+TEST_WORKSPACE_SRCS = tests/test_workspace.c embed.c embed_config.c qwen_tokenizer.c \
qwen_safetensors.c $(KERNEL_SRCS)
-TEST_LATE_SRCS = tests/test_late.c embed.c qwen_tokenizer.c \
+TEST_LATE_SRCS = tests/test_late.c embed.c embed_config.c qwen_tokenizer.c \
qwen_safetensors.c $(KERNEL_SRCS)
-TEST_SERVER_SRCS = tests/test_server.c embed.c \
+TEST_SERVER_SRCS = tests/test_server.c embed.c embed_config.c \
qwen_tokenizer.c qwen_safetensors.c $(KERNEL_SRCS) \
deps/ae/ae.c deps/ae/anet.c deps/ae/monotonic.c
# The CLI check builds the real CLI (CPU backend) plus a driver that runs it.
-TEST_CLI_BIN_SRCS = embed_cli.c embed.c \
+TEST_CLI_BIN_SRCS = embed_cli.c embed.c embed_config.c \
qwen_tokenizer.c qwen_safetensors.c $(KERNEL_SRCS)
test:
@@ -322,7 +323,7 @@ bench:
bench-model:
@test -n "$(MODEL_DIR)" || { echo "usage: make bench-model MODEL_DIR=/path/to/model-dir"; exit 1; }
$(CC) $(BENCH_CC_FLAGS) -o bench/bench_model \
- bench/bench_model.c embed.c qwen_safetensors.c $(KERNEL_SRCS) \
+ bench/bench_model.c embed.c embed_config.c qwen_safetensors.c $(KERNEL_SRCS) \
-lm -lpthread $(TEST_BLAS_LDFLAGS)
@mkdir -p bench/results
./bench/bench_model "$(MODEL_DIR)" --json bench/results/model-$(BENCH_STAMP).json
@@ -353,13 +354,13 @@ $(SERVER_TARGET): $(LIB) embed_server.o
# =============================================================================
# Compile rules
# =============================================================================
-embed.o: embed.c embed.h qwen_kernels.h qwen_safetensors.h
+embed.o: embed.c embed.h embed_config.h qwen_kernels.h qwen_safetensors.h
$(CC) $(CFLAGS) -Ideps/ae -c -o $@ $<
embed_server.o: embed_server.c embed_server.h embed_build.h embed.h embed_mlx.h qwen_tokenizer.h deps/ae/ae.h deps/ae/anet.h
$(CC) $(CFLAGS) $(VERSION_CFLAGS) $(CJSON_CFLAGS) -Ideps/ae -c -o $@ $<
-embed_mlx.o: embed_mlx.c embed_mlx.h embed.h qwen_safetensors.h
+embed_mlx.o: embed_mlx.c embed_mlx.h embed_config.h embed.h qwen_safetensors.h
$(CC) $(CFLAGS) -c -o $@ $<
embed_cli.o: embed_cli.c embed_build.h embed.h qwen_kernels.h qwen_tokenizer.h
@@ -384,6 +385,9 @@ qwen_tokenizer.o: qwen_tokenizer.c qwen_tokenizer.h qwen_kernels.h
qwen_safetensors.o: qwen_safetensors.c qwen_safetensors.h
$(CC) $(CFLAGS) -c -o $@ $<
+embed_config.o: embed_config.c embed_config.h embed.h
+ $(CC) $(CFLAGS) -c -o $@ $<
+
deps/ae/%.o: deps/ae/%.c deps/ae/ae.h deps/ae/anet.h deps/ae/monotonic.h
$(CC) $(CFLAGS) -Ideps/ae -c -o $@ $<
diff --git a/embed.c b/embed.c
index 67545a8..c05adfc 100644
--- a/embed.c
+++ b/embed.c
@@ -1,4 +1,5 @@
#include "embed.h"
+#include "embed_config.h"
#include "embed_internal.h"
#include "qwen_kernels.h"
#include "qwen_safetensors.h"
@@ -148,151 +149,6 @@ static int realloc_ints(int **ptr, size_t count) {
}
/* ========================================================================
- * Minimal JSON helpers for config.json parsing
- * ======================================================================== */
-
-static const char *skip_ws(const char *p) {
- while (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
- p++;
- return p;
-}
-
-/* Find "key": <value> in a flat JSON object. Returns pointer to value start. */
-static const char *json_find_key(const char *json, const char *key) {
- size_t klen = strlen(key);
- const char *p = json;
- while ((p = strstr(p, key)) != NULL) {
- /* Check it's actually a quoted key: "key" */
- if (p > json && *(p - 1) == '"') {
- const char *after = p + klen;
- if (*after == '"') {
- after = skip_ws(after + 1);
- if (*after == ':')
- return skip_ws(after + 1);
- }
- }
- p += klen;
- }
- return NULL;
-}
-
-static int json_get_int(const char *json, const char *key, int fallback) {
- const char *v = json_find_key(json, key);
- if (!v)
- return fallback;
- return atoi(v);
-}
-
-static double json_get_double(const char *json, const char *key, double fallback) {
- const char *v = json_find_key(json, key);
- if (!v)
- return fallback;
- return atof(v);
-}
-
-static int json_string_equals(const char *json, const char *key, const char *expected) {
- const char *v = json_find_key(json, key);
- if (!v || *v != '"')
- return 0;
- v++;
- size_t len = strlen(expected);
- return strncmp(v, expected, len) == 0 && v[len] == '"';
-}
-
-static int parse_config(embed_config_t *cfg, const char *model_dir) {
- char path[1024];
- int n = snprintf(path, sizeof(path), "%s/config.json", model_dir);
- if (n < 0 || (size_t)n >= sizeof(path)) {
- fprintf(stderr, "embed_model_load: model path too long\n");
- return -1;
- }
-
- FILE *f = fopen(path, "rb");
- if (!f) {
- fprintf(stderr, "embed_model_load: cannot open %s\n", path);
- return -1;
- }
-
- if (fseek(f, 0, SEEK_END) != 0) {
- fclose(f);
- return -1;
- }
- long sz = ftell(f);
- if (sz < 0 || (size_t)sz == SIZE_MAX || fseek(f, 0, SEEK_SET) != 0) {
- fclose(f);
- return -1;
- }
- char *buf = (char *)malloc((size_t)sz + 1);
- if (!buf || fread(buf, 1, (size_t)sz, f) != (size_t)sz) {
- fclose(f);
- free(buf);
- return -1;
- }
- fclose(f);
- buf[sz] = '\0';
-
- cfg->hidden_size = json_get_int(buf, "hidden_size", 0);
- cfg->n_layers = json_get_int(buf, "num_hidden_layers", 0);
- cfg->n_heads = json_get_int(buf, "num_attention_heads", 0);
- cfg->n_kv_heads = json_get_int(buf, "num_key_value_heads", 0);
- cfg->head_dim = json_get_int(buf, "head_dim", EMBED_HEAD_DIM);
- cfg->intermediate_size = json_get_int(buf, "intermediate_size", 0);
- cfg->vocab_size = json_get_int(buf, "vocab_size", EMBED_VOCAB_SIZE);
- cfg->rms_norm_eps = (float)json_get_double(buf, "rms_norm_eps", 1e-6);
- cfg->rope_theta = (float)json_get_double(buf, "rope_theta", 1000000.0);
- cfg->attention_mode = EMBED_ATTENTION_BIDIRECTIONAL;
- cfg->pooling_mode = EMBED_POOL_MEAN;
- cfg->normalize_embeddings = 0;
- cfg->append_terminal_token = 0;
-
- if (json_string_equals(buf, "model_type", "qwen3")) {
- cfg->attention_mode = EMBED_ATTENTION_CAUSAL;
- cfg->pooling_mode = EMBED_POOL_LAST_TOKEN;
- cfg->normalize_embeddings = 1;
- cfg->append_terminal_token = 1;
- }
-
- cfg->q_dim = 0;
- cfg->kv_dim = 0;
- if (cfg->n_heads > 0 && cfg->head_dim > 0 && cfg->n_heads <= INT_MAX / cfg->head_dim)
- cfg->q_dim = cfg->n_heads * cfg->head_dim;
- if (cfg->n_kv_heads > 0 && cfg->head_dim > 0 && cfg->n_kv_heads <= INT_MAX / cfg->head_dim)
- cfg->kv_dim = cfg->n_kv_heads * cfg->head_dim;
-
- free(buf);
-
- /* Sanity checks */
- if (cfg->hidden_size <= 0 || cfg->n_layers <= 0 || cfg->n_heads <= 0 || cfg->n_kv_heads <= 0 ||
- cfg->head_dim <= 0 || cfg->intermediate_size <= 0 || cfg->vocab_size <= 0 ||
- cfg->q_dim <= 0 || cfg->kv_dim <= 0 || (cfg->head_dim & 1) ||
- cfg->n_heads % cfg->n_kv_heads != 0 || !isfinite(cfg->rms_norm_eps) ||
- cfg->rms_norm_eps <= 0.0f || !isfinite(cfg->rope_theta) || cfg->rope_theta <= 0.0f) {
- fprintf(stderr,
- "embed_model_load: invalid config in %s "
- "(hidden=%d, layers=%d, heads=%d/%d, head_dim=%d, inter=%d)\n",
- path, cfg->hidden_size, cfg->n_layers, cfg->n_heads, cfg->n_kv_heads, cfg->head_dim,
- cfg->intermediate_size);
- return -1;
- }
- if (cfg->n_layers > EMBED_MAX_LAYERS) {
- fprintf(stderr, "embed_model_load: too many layers (%d > %d)\n", cfg->n_layers,
- EMBED_MAX_LAYERS);
- return -1;
- }
-
- if (embed_verbose >= 1)
- fprintf(stderr,
- "config: hidden=%d, layers=%d, heads=%d/%d, "
- "inter=%d, head_dim=%d, attention=%s, pooling=%s\n",
- cfg->hidden_size, cfg->n_layers, cfg->n_heads, cfg->n_kv_heads,
- cfg->intermediate_size, cfg->head_dim,
- cfg->attention_mode == EMBED_ATTENTION_CAUSAL ? "causal" : "bidirectional",
- cfg->pooling_mode == EMBED_POOL_LAST_TOKEN ? "last-token" : "mean");
-
- return 0;
-}
-
-/* ========================================================================
* Weight loading (direct mmap pointers into safetensors)
* ======================================================================== */
@@ -636,7 +492,7 @@ model_load_range_ex(const char *model_dir, int layer_start, int layer_end, int a
return NULL;
/* Parse config.json */
- if (parse_config(&model->config, model_dir) != 0) {
+ if (embed_config_parse(&model->config, model_dir) != 0) {
free(model);
return NULL;
}
diff --git a/embed_config.c b/embed_config.c
new file mode 100644
index 0000000..5bbfe48
--- /dev/null
+++ b/embed_config.c
@@ -0,0 +1,141 @@
+#include "embed_config.h"
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+static const char *skip_ws(const char *p) {
+ while (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
+ p++;
+ return p;
+}
+
+static const char *json_find_key(const char *json, const char *key) {
+ size_t klen = strlen(key);
+ const char *p = json;
+ while ((p = strstr(p, key)) != NULL) {
+ if (p > json && p[-1] == '"') {
+ const char *after = p + klen;
+ if (*after == '"') {
+ after = skip_ws(after + 1);
+ if (*after == ':')
+ return skip_ws(after + 1);
+ }
+ }
+ p += klen;
+ }
+ return NULL;
+}
+
+static int json_get_int(const char *json, const char *key, int fallback) {
+ const char *v = json_find_key(json, key);
+ return v ? atoi(v) : fallback;
+}
+
+static double json_get_double(const char *json, const char *key, double fallback) {
+ const char *v = json_find_key(json, key);
+ return v ? atof(v) : fallback;
+}
+
+static int json_string_equals(const char *json, const char *key, const char *expected) {
+ const char *v = json_find_key(json, key);
+ if (!v || *v != '"')
+ return 0;
+ v++;
+ size_t len = strlen(expected);
+ return strncmp(v, expected, len) == 0 && v[len] == '"';
+}
+
+int embed_config_parse(embed_config_t *cfg, const char *model_dir) {
+ char path[1024];
+ int n = snprintf(path, sizeof(path), "%s/config.json", model_dir);
+ if (n < 0 || (size_t)n >= sizeof(path)) {
+ fprintf(stderr, "embed_config: model path too long\n");
+ return -1;
+ }
+
+ FILE *f = fopen(path, "rb");
+ if (!f) {
+ fprintf(stderr, "embed_config: cannot open %s\n", path);
+ return -1;
+ }
+
+ if (fseek(f, 0, SEEK_END) != 0) {
+ fclose(f);
+ return -1;
+ }
+ long sz = ftell(f);
+ if (sz < 0 || (size_t)sz == SIZE_MAX || fseek(f, 0, SEEK_SET) != 0) {
+ fclose(f);
+ return -1;
+ }
+ char *buf = (char *)malloc((size_t)sz + 1);
+ if (!buf || fread(buf, 1, (size_t)sz, f) != (size_t)sz) {
+ fclose(f);
+ free(buf);
+ return -1;
+ }
+ fclose(f);
+ buf[sz] = '\0';
+
+ cfg->hidden_size = json_get_int(buf, "hidden_size", 0);
+ cfg->n_layers = json_get_int(buf, "num_hidden_layers", 0);
+ cfg->n_heads = json_get_int(buf, "num_attention_heads", 0);
+ cfg->n_kv_heads = json_get_int(buf, "num_key_value_heads", 0);
+ cfg->head_dim = json_get_int(buf, "head_dim", EMBED_HEAD_DIM);
+ cfg->intermediate_size = json_get_int(buf, "intermediate_size", 0);
+ cfg->vocab_size = json_get_int(buf, "vocab_size", EMBED_VOCAB_SIZE);
+ cfg->rms_norm_eps = (float)json_get_double(buf, "rms_norm_eps", 1e-6);
+ cfg->rope_theta = (float)json_get_double(buf, "rope_theta", 1000000.0);
+ cfg->attention_mode = EMBED_ATTENTION_BIDIRECTIONAL;
+ cfg->pooling_mode = EMBED_POOL_MEAN;
+ cfg->normalize_embeddings = 0;
+ cfg->append_terminal_token = 0;
+
+ if (json_string_equals(buf, "model_type", "qwen3")) {
+ cfg->attention_mode = EMBED_ATTENTION_CAUSAL;
+ cfg->pooling_mode = EMBED_POOL_LAST_TOKEN;
+ cfg->normalize_embeddings = 1;
+ cfg->append_terminal_token = 1;
+ }
+
+ cfg->q_dim = 0;
+ cfg->kv_dim = 0;
+ if (cfg->n_heads > 0 && cfg->head_dim > 0 && cfg->n_heads <= INT_MAX / cfg->head_dim)
+ cfg->q_dim = cfg->n_heads * cfg->head_dim;
+ if (cfg->n_kv_heads > 0 && cfg->head_dim > 0 && cfg->n_kv_heads <= INT_MAX / cfg->head_dim)
+ cfg->kv_dim = cfg->n_kv_heads * cfg->head_dim;
+
+ free(buf);
+
+ if (cfg->hidden_size <= 0 || cfg->n_layers <= 0 || cfg->n_heads <= 0 || cfg->n_kv_heads <= 0 ||
+ cfg->head_dim <= 0 || cfg->intermediate_size <= 0 || cfg->vocab_size <= 0 ||
+ cfg->q_dim <= 0 || cfg->kv_dim <= 0 || (cfg->head_dim & 1) ||
+ cfg->n_heads % cfg->n_kv_heads != 0 || !isfinite(cfg->rms_norm_eps) ||
+ cfg->rms_norm_eps <= 0.0f || !isfinite(cfg->rope_theta) || cfg->rope_theta <= 0.0f) {
+ fprintf(stderr,
+ "embed_config: invalid config in %s "
+ "(hidden=%d, layers=%d, heads=%d/%d, head_dim=%d, inter=%d)\n",
+ path, cfg->hidden_size, cfg->n_layers, cfg->n_heads, cfg->n_kv_heads, cfg->head_dim,
+ cfg->intermediate_size);
+ return -1;
+ }
+ if (cfg->n_layers > EMBED_MAX_LAYERS) {
+ fprintf(stderr, "embed_config: too many layers (%d > %d)\n", cfg->n_layers,
+ EMBED_MAX_LAYERS);
+ return -1;
+ }
+
+ if (embed_verbose >= 1)
+ fprintf(stderr,
+ "config: hidden=%d, layers=%d, heads=%d/%d, "
+ "inter=%d, head_dim=%d, attention=%s, pooling=%s\n",
+ cfg->hidden_size, cfg->n_layers, cfg->n_heads, cfg->n_kv_heads,
+ cfg->intermediate_size, cfg->head_dim,
+ cfg->attention_mode == EMBED_ATTENTION_CAUSAL ? "causal" : "bidirectional",
+ cfg->pooling_mode == EMBED_POOL_LAST_TOKEN ? "last-token" : "mean");
+
+ return 0;
+}
diff --git a/embed_config.h b/embed_config.h
new file mode 100644
index 0000000..e8353d3
--- /dev/null
+++ b/embed_config.h
@@ -0,0 +1,8 @@
+#ifndef EMBED_CONFIG_H
+#define EMBED_CONFIG_H
+
+#include "embed.h"
+
+int embed_config_parse(embed_config_t *cfg, const char *model_dir);
+
+#endif /* EMBED_CONFIG_H */
diff --git a/embed_mlx.c b/embed_mlx.c
index b9b2137..de82f83 100644
--- a/embed_mlx.c
+++ b/embed_mlx.c
@@ -1,4 +1,5 @@
#include "embed_mlx.h"
+#include "embed_config.h"
#include "embed.h"
#include "qwen_safetensors.h"
@@ -322,130 +323,6 @@ static void upcast_f32_inplace(mlx_array *a, mlx_stream s) {
* Load
* ======================================================================== */
-/* We duplicate a minimal config parser here so MLX can load standalone.
- * (Same logic as embed.c:parse_config, kept inline for simplicity.) */
-
-static const char *skip_ws_m(const char *p) {
- while (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
- p++;
- return p;
-}
-
-static const char *json_find_m(const char *json, const char *key) {
- size_t klen = strlen(key);
- const char *p = json;
- while ((p = strstr(p, key)) != NULL) {
- if (p > json && *(p - 1) == '"') {
- const char *a = p + klen;
- if (*a == '"') {
- a = skip_ws_m(a + 1);
- if (*a == ':')
- return skip_ws_m(a + 1);
- }
- }
- p += klen;
- }
- return NULL;
-}
-
-static int json_string_equals_m(const char *json, const char *key, const char *expected) {
- const char *v = json_find_m(json, key);
- if (!v || *v != '"')
- return 0;
- v++;
- size_t len = strlen(expected);
- return strncmp(v, expected, len) == 0 && v[len] == '"';
-}
-
-static int mlx_parse_config(embed_config_t *cfg, const char *model_dir) {
- char path[1024];
- int path_len = snprintf(path, sizeof(path), "%s/config.json", model_dir);
- if (path_len < 0 || (size_t)path_len >= sizeof(path)) {
- fprintf(stderr, "mlx: model path too long\n");
- return -1;
- }
- FILE *f = fopen(path, "rb");
- if (!f) {
- fprintf(stderr, "mlx: cannot open %s\n", path);
- return -1;
- }
- if (fseek(f, 0, SEEK_END) != 0) {
- fclose(f);
- return -1;
- }
- long sz = ftell(f);
- if (sz < 0 || (size_t)sz == SIZE_MAX || fseek(f, 0, SEEK_SET) != 0) {
- fclose(f);
- return -1;
- }
- char *buf = (char *)malloc((size_t)sz + 1);
- if (!buf || fread(buf, 1, (size_t)sz, f) != (size_t)sz) {
- fclose(f);
- free(buf);
- return -1;
- }
- fclose(f);
- buf[sz] = '\0';
-
-#define GI(k, fb, def) \
- do { \
- const char *v = json_find_m(buf, k); \
- cfg->fb = v ? atoi(v) : def; \
- } while (0)
-#define GF(k, fb, def) \
- do { \
- const char *v = json_find_m(buf, k); \
- cfg->fb = v ? (float)atof(v) : def; \
- } while (0)
- GI("hidden_size", hidden_size, 0);
- GI("num_hidden_layers", n_layers, 0);
- GI("num_attention_heads", n_heads, 0);
- GI("num_key_value_heads", n_kv_heads, 0);
- GI("head_dim", head_dim, EMBED_HEAD_DIM);
- GI("intermediate_size", intermediate_size, 0);
- GI("vocab_size", vocab_size, EMBED_VOCAB_SIZE);
- GF("rms_norm_eps", rms_norm_eps, 1e-6f);
- GF("rope_theta", rope_theta, 1000000.0f);
-#undef GI
-#undef GF
- cfg->attention_mode = EMBED_ATTENTION_BIDIRECTIONAL;
- cfg->pooling_mode = EMBED_POOL_MEAN;
- cfg->normalize_embeddings = 0;
- cfg->append_terminal_token = 0;
- if (json_string_equals_m(buf, "model_type", "qwen3")) {
- cfg->attention_mode = EMBED_ATTENTION_CAUSAL;
- cfg->pooling_mode = EMBED_POOL_LAST_TOKEN;
- cfg->normalize_embeddings = 1;
- cfg->append_terminal_token = 1;
- }
-
- cfg->q_dim = 0;
- cfg->kv_dim = 0;
- if (cfg->n_heads > 0 && cfg->head_dim > 0 && cfg->n_heads <= INT_MAX / cfg->head_dim)
- cfg->q_dim = cfg->n_heads * cfg->head_dim;
- if (cfg->n_kv_heads > 0 && cfg->head_dim > 0 && cfg->n_kv_heads <= INT_MAX / cfg->head_dim)
- cfg->kv_dim = cfg->n_kv_heads * cfg->head_dim;
- free(buf);
-
- if (cfg->hidden_size <= 0 || cfg->n_layers <= 0 || cfg->n_heads <= 0 || cfg->n_kv_heads <= 0 ||
- cfg->head_dim <= 0 || cfg->intermediate_size <= 0 || cfg->vocab_size <= 0 ||
- cfg->q_dim <= 0 || cfg->kv_dim <= 0 || (cfg->head_dim & 1) ||
- cfg->n_heads % cfg->n_kv_heads != 0 || !isfinite(cfg->rms_norm_eps) ||
- cfg->rms_norm_eps <= 0.0f || !isfinite(cfg->rope_theta) || cfg->rope_theta <= 0.0f) {
- fprintf(stderr,
- "mlx: invalid config in %s "
- "(hidden=%d, layers=%d, heads=%d/%d, head_dim=%d, inter=%d)\n",
- path, cfg->hidden_size, cfg->n_layers, cfg->n_heads, cfg->n_kv_heads, cfg->head_dim,
- cfg->intermediate_size);
- return -1;
- }
- if (cfg->n_layers > EMBED_MAX_LAYERS) {
- fprintf(stderr, "mlx: too many layers (%d > %d)\n", cfg->n_layers, EMBED_MAX_LAYERS);
- return -1;
- }
- return 0;
-}
-
static int mlx_model_dir_has_file(const char *model_dir, const char *rel) {
char path[1024];
int n = snprintf(path, sizeof(path), "%s/%s", model_dir, rel);
@@ -516,7 +393,7 @@ static embed_mlx_ctx_t *mlx_load_range_ex(const char *model_dir,
ctx->quantize_bits = options.quantize_bits;
ctx->quantize_group_size = options.quantize_group_size;
- if (mlx_parse_config(&ctx->config, model_dir) != 0) {
+ if (embed_config_parse(&ctx->config, model_dir) != 0) {
free(ctx);
return NULL;
}