diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-15 11:45:26 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-15 11:45:26 +0200 |
| commit | 588feec258dd309e4b947c25bfacad7ad956547f (patch) | |
| tree | 98eca640f17e9815748358fba41982806259b281 | |
| parent | d7b62df9dfae32c62c2d7ec3da79ab3239af8d31 (diff) | |
| download | ffwd-588feec258dd309e4b947c25bfacad7ad956547f.tar.gz ffwd-588feec258dd309e4b947c25bfacad7ad956547f.zip | |
api: accept Hugging Face model weight prefix
| -rw-r--r-- | embed.c | 66 | ||||
| -rw-r--r-- | embed_mlx.c | 43 | ||||
| -rw-r--r-- | qwen_safetensors.c | 12 | ||||
| -rw-r--r-- | qwen_safetensors.h | 4 | ||||
| -rw-r--r-- | tests/test_bf16_model.c | 36 | ||||
| -rw-r--r-- | tests/tiny_model.h | 38 |
6 files changed, 134 insertions, 65 deletions
@@ -667,14 +667,24 @@ model_load_range_ex(const char *model_dir, int layer_start, int layer_end, int a } model->safetensors = ms; + const char *weight_prefix = multi_safetensors_weight_prefix(ms, "embed_tokens.weight"); + if (!weight_prefix) { + fprintf(stderr, "embed_model_load: neither embed_tokens.weight nor " + "model.embed_tokens.weight was found\n"); + goto fail; + } + if (embed_verbose >= 1) fprintf(stderr, "embed_model_load: loading weights from %s\n", model_dir); + char name[256]; + /* Only the first stage needs token embeddings. */ embed_weights_t *w = &model->weights; if (layer_start == 0) { const int64_t embed_shape[2] = {cfg->vocab_size, cfg->hidden_size}; - w->embed_tokens = load_weight_direct(ms, "embed_tokens.weight", embed_shape, 2); + snprintf(name, sizeof(name), "%sembed_tokens.weight", weight_prefix); + w->embed_tokens = load_weight_direct(ms, name, embed_shape, 2); if (!w->embed_tokens.data) goto fail; } @@ -685,40 +695,39 @@ model_load_range_ex(const char *model_dir, int layer_start, int layer_end, int a goto fail; /* Per-layer weights */ - char name[256]; for (int i = layer_start; i < layer_end; i++) { embed_layer_t *l = &w->layers[i]; -#define LOAD_NORM(field, fmt, d0) \ - do { \ - const int64_t expect[1] = {(d0)}; \ - snprintf(name, sizeof(name), fmt, i); \ - l->field = load_norm_f32(ms, name, expect, 1); \ - if (!l->field) \ - goto fail; \ +#define LOAD_NORM(field, fmt, d0) \ + do { \ + const int64_t expect[1] = {(d0)}; \ + snprintf(name, sizeof(name), fmt, weight_prefix, i); \ + l->field = load_norm_f32(ms, name, expect, 1); \ + if (!l->field) \ + goto fail; \ } while (0) -#define LOAD2(field, fmt, d0, d1) \ - do { \ - const int64_t expect[2] = {(d0), (d1)}; \ - snprintf(name, sizeof(name), fmt, i); \ - l->field = load_weight_direct(ms, name, expect, 2); \ - if (!l->field.data) \ - goto fail; \ +#define LOAD2(field, fmt, d0, d1) \ + do { \ + const int64_t expect[2] = {(d0), (d1)}; \ + snprintf(name, sizeof(name), fmt, weight_prefix, i); \ + l->field = load_weight_direct(ms, name, expect, 2); \ + if (!l->field.data) \ + goto fail; \ } while (0) - LOAD2(wq, "layers.%d.self_attn.q_proj.weight", cfg->q_dim, cfg->hidden_size); - LOAD2(wk, "layers.%d.self_attn.k_proj.weight", cfg->kv_dim, cfg->hidden_size); - LOAD2(wv, "layers.%d.self_attn.v_proj.weight", cfg->kv_dim, cfg->hidden_size); - LOAD2(wo, "layers.%d.self_attn.o_proj.weight", cfg->hidden_size, cfg->q_dim); - LOAD_NORM(q_norm, "layers.%d.self_attn.q_norm.weight", cfg->head_dim); - LOAD_NORM(k_norm, "layers.%d.self_attn.k_norm.weight", cfg->head_dim); - LOAD_NORM(input_norm, "layers.%d.input_layernorm.weight", cfg->hidden_size); - LOAD_NORM(post_attn_norm, "layers.%d.post_attention_layernorm.weight", cfg->hidden_size); - LOAD2(gate_proj, "layers.%d.mlp.gate_proj.weight", cfg->intermediate_size, + LOAD2(wq, "%slayers.%d.self_attn.q_proj.weight", cfg->q_dim, cfg->hidden_size); + LOAD2(wk, "%slayers.%d.self_attn.k_proj.weight", cfg->kv_dim, cfg->hidden_size); + LOAD2(wv, "%slayers.%d.self_attn.v_proj.weight", cfg->kv_dim, cfg->hidden_size); + LOAD2(wo, "%slayers.%d.self_attn.o_proj.weight", cfg->hidden_size, cfg->q_dim); + LOAD_NORM(q_norm, "%slayers.%d.self_attn.q_norm.weight", cfg->head_dim); + LOAD_NORM(k_norm, "%slayers.%d.self_attn.k_norm.weight", cfg->head_dim); + LOAD_NORM(input_norm, "%slayers.%d.input_layernorm.weight", cfg->hidden_size); + LOAD_NORM(post_attn_norm, "%slayers.%d.post_attention_layernorm.weight", cfg->hidden_size); + LOAD2(gate_proj, "%slayers.%d.mlp.gate_proj.weight", cfg->intermediate_size, cfg->hidden_size); - LOAD2(up_proj, "layers.%d.mlp.up_proj.weight", cfg->intermediate_size, cfg->hidden_size); - LOAD2(down_proj, "layers.%d.mlp.down_proj.weight", cfg->hidden_size, + LOAD2(up_proj, "%slayers.%d.mlp.up_proj.weight", cfg->intermediate_size, cfg->hidden_size); + LOAD2(down_proj, "%slayers.%d.mlp.down_proj.weight", cfg->hidden_size, cfg->intermediate_size); #undef LOAD2 @@ -731,7 +740,8 @@ model_load_range_ex(const char *model_dir, int layer_start, int layer_end, int a /* Only the final stage applies the output RMSNorm. */ if (layer_end == cfg->n_layers) { const int64_t norm_shape[1] = {cfg->hidden_size}; - w->norm = load_norm_f32(ms, "norm.weight", norm_shape, 1); + snprintf(name, sizeof(name), "%snorm.weight", weight_prefix); + w->norm = load_norm_f32(ms, name, norm_shape, 1); if (!w->norm) goto fail; } diff --git a/embed_mlx.c b/embed_mlx.c index e18a0a4..b9b2137 100644 --- a/embed_mlx.c +++ b/embed_mlx.c @@ -540,13 +540,22 @@ static embed_mlx_ctx_t *mlx_load_range_ex(const char *model_dir, return NULL; } + const char *weight_prefix = multi_safetensors_weight_prefix(ctx->ms, "embed_tokens.weight"); + if (!weight_prefix) { + fprintf(stderr, "mlx: neither embed_tokens.weight nor " + "model.embed_tokens.weight was found\n"); + goto fail; + } + int h = c->hidden_size, qd = c->q_dim, kvd = c->kv_dim; int inter = c->intermediate_size, hd = c->head_dim; + char name[256]; /* Only the first stage needs token embeddings. */ if (layer_start == 0) { int emb_shape[] = {c->vocab_size, h}; - ctx->embed_tokens = load_tensor(ctx->ms, "embed_tokens.weight", emb_shape, 2); + snprintf(name, sizeof(name), "%sembed_tokens.weight", weight_prefix); + ctx->embed_tokens = load_tensor(ctx->ms, name, emb_shape, 2); if (!arr_ok(ctx->embed_tokens)) goto fail; } @@ -556,52 +565,51 @@ static embed_mlx_ctx_t *mlx_load_range_ex(const char *model_dir, if (!ctx->layers) goto fail; - char name[256]; for (int i = layer_start; i < layer_end; i++) { mlx_layer_t *l = &ctx->layers[i]; #define LD(fld, fmt, ...) \ do { \ - snprintf(name, sizeof(name), fmt, i); \ + snprintf(name, sizeof(name), fmt, weight_prefix, i); \ int sh[] = {__VA_ARGS__}; \ l->fld = load_linear_tensor(ctx->ms, name, sh, sizeof(sh) / sizeof(sh[0])); \ if (!linear_ok(l->fld)) \ goto fail; \ } while (0) - LD(wq, "layers.%d.self_attn.q_proj.weight", qd, h); - LD(wk, "layers.%d.self_attn.k_proj.weight", kvd, h); - LD(wv, "layers.%d.self_attn.v_proj.weight", kvd, h); - LD(wo, "layers.%d.self_attn.o_proj.weight", h, qd); + LD(wq, "%slayers.%d.self_attn.q_proj.weight", qd, h); + LD(wk, "%slayers.%d.self_attn.k_proj.weight", kvd, h); + LD(wv, "%slayers.%d.self_attn.v_proj.weight", kvd, h); + LD(wo, "%slayers.%d.self_attn.o_proj.weight", h, qd); #undef LD #define LD(fld, fmt, ...) \ do { \ - snprintf(name, sizeof(name), fmt, i); \ + snprintf(name, sizeof(name), fmt, weight_prefix, i); \ int sh[] = {__VA_ARGS__}; \ l->fld = load_tensor(ctx->ms, name, sh, sizeof(sh) / sizeof(sh[0])); \ if (!arr_ok(l->fld)) \ goto fail; \ } while (0) - LD(q_norm, "layers.%d.self_attn.q_norm.weight", hd); - LD(k_norm, "layers.%d.self_attn.k_norm.weight", hd); - LD(input_norm, "layers.%d.input_layernorm.weight", h); - LD(post_attn_norm, "layers.%d.post_attention_layernorm.weight", h); + LD(q_norm, "%slayers.%d.self_attn.q_norm.weight", hd); + LD(k_norm, "%slayers.%d.self_attn.k_norm.weight", hd); + LD(input_norm, "%slayers.%d.input_layernorm.weight", h); + LD(post_attn_norm, "%slayers.%d.post_attention_layernorm.weight", h); #undef LD #define LD(fld, fmt, ...) \ do { \ - snprintf(name, sizeof(name), fmt, i); \ + snprintf(name, sizeof(name), fmt, weight_prefix, i); \ int sh[] = {__VA_ARGS__}; \ l->fld = load_linear_tensor(ctx->ms, name, sh, sizeof(sh) / sizeof(sh[0])); \ if (!linear_ok(l->fld)) \ goto fail; \ } while (0) - LD(gate_proj, "layers.%d.mlp.gate_proj.weight", inter, h); - LD(up_proj, "layers.%d.mlp.up_proj.weight", inter, h); - LD(down_proj, "layers.%d.mlp.down_proj.weight", h, inter); + LD(gate_proj, "%slayers.%d.mlp.gate_proj.weight", inter, h); + LD(up_proj, "%slayers.%d.mlp.up_proj.weight", inter, h); + LD(down_proj, "%slayers.%d.mlp.down_proj.weight", h, inter); #undef LD if (embed_verbose >= 2) @@ -610,7 +618,8 @@ static embed_mlx_ctx_t *mlx_load_range_ex(const char *model_dir, if (layer_end == c->n_layers) { int ns[] = {h}; - ctx->norm = load_tensor(ctx->ms, "norm.weight", ns, 1); + snprintf(name, sizeof(name), "%snorm.weight", weight_prefix); + ctx->norm = load_tensor(ctx->ms, name, ns, 1); if (!arr_ok(ctx->norm)) goto fail; } diff --git a/qwen_safetensors.c b/qwen_safetensors.c index e1c0e87..2222a2a 100644 --- a/qwen_safetensors.c +++ b/qwen_safetensors.c @@ -545,3 +545,15 @@ const safetensor_t *multi_safetensors_find(const multi_safetensors_t *ms, } return NULL; } + +const char *multi_safetensors_weight_prefix(const multi_safetensors_t *ms, + const char *tensor_name) { + if (multi_safetensors_find(ms, tensor_name, NULL)) + return ""; + + char prefixed[sizeof(((safetensor_t *)0)->name)]; + int n = snprintf(prefixed, sizeof(prefixed), "model.%s", tensor_name); + if (n < 0 || (size_t)n >= sizeof(prefixed)) + return NULL; + return multi_safetensors_find(ms, prefixed, NULL) ? "model." : NULL; +} diff --git a/qwen_safetensors.h b/qwen_safetensors.h index 5e839a9..c060847 100644 --- a/qwen_safetensors.h +++ b/qwen_safetensors.h @@ -61,6 +61,10 @@ const safetensor_t *multi_safetensors_find(const multi_safetensors_t *ms, const char *name, safetensors_file_t **out_sf); +/* Resolve an optional Hugging Face "model." namespace for a backbone tensor. + * Returns "", "model.", or NULL when neither form exists. */ +const char *multi_safetensors_weight_prefix(const multi_safetensors_t *ms, const char *tensor_name); + /* Get raw pointer to tensor data (within mmap'd region) */ const void *safetensors_data(const safetensors_file_t *sf, const safetensor_t *t); diff --git a/tests/test_bf16_model.c b/tests/test_bf16_model.c index 9a5d842..48e4766 100644 --- a/tests/test_bf16_model.c +++ b/tests/test_bf16_model.c @@ -12,7 +12,7 @@ #include <unistd.h> int main(void) { - char root[1024], f32_dir[1088], bf16_dir[1088]; + char root[1024], f32_dir[1088], bf16_dir[1088], prefixed_dir[1088]; snprintf(root, sizeof(root), "%s/embed-bf16-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); if (!mkdtemp(root)) { @@ -21,8 +21,11 @@ int main(void) { } snprintf(f32_dir, sizeof(f32_dir), "%s/f32", root); snprintf(bf16_dir, sizeof(bf16_dir), "%s/bf16", root); - if (mkdir(f32_dir, 0755) != 0 || mkdir(bf16_dir, 0755) != 0 || - tm_write_model(f32_dir, "F32") != 0 || tm_write_model(bf16_dir, "BF16") != 0) { + snprintf(prefixed_dir, sizeof(prefixed_dir), "%s/prefixed", root); + static const tm_dims_t dims = {4, 2, 1, 2, 8, 16}; + if (mkdir(f32_dir, 0755) != 0 || mkdir(bf16_dir, 0755) != 0 || mkdir(prefixed_dir, 0755) != 0 || + tm_write_model(f32_dir, "F32") != 0 || tm_write_model(bf16_dir, "BF16") != 0 || + tm_write_prefixed_model_dims(prefixed_dir, "F32", &dims) != 0) { fprintf(stderr, "fixture creation failed\n"); return 2; } @@ -30,7 +33,8 @@ int main(void) { int ids[] = {1, 2, 3, 4}; embed_model_t *mf32 = embed_model_load(f32_dir); embed_model_t *mbf16 = embed_model_load(bf16_dir); - if (!mf32 || !mbf16) { + embed_model_t *mprefixed = embed_model_load(prefixed_dir); + if (!mf32 || !mbf16 || !mprefixed) { fprintf(stderr, "model load failed\n"); return 1; } @@ -39,36 +43,44 @@ int main(void) { int dim = cfg->hidden_size; float *a = (float *)malloc((size_t)dim * sizeof(float)); float *b = (float *)malloc((size_t)dim * sizeof(float)); + float *c = (float *)malloc((size_t)dim * sizeof(float)); embed_workspace_t *wf32 = embed_workspace_new(mf32); embed_workspace_t *wbf16 = embed_workspace_new(mbf16); - if (!a || !b || !wf32 || !wbf16) { + embed_workspace_t *wprefixed = embed_workspace_new(mprefixed); + if (!a || !b || !c || !wf32 || !wbf16 || !wprefixed) { fprintf(stderr, "allocation failed\n"); return 1; } if (embed_model_encode_into(mf32, wf32, ids, 4, a) != 0 || - embed_model_encode_into(mbf16, wbf16, ids, 4, b) != 0) { + embed_model_encode_into(mbf16, wbf16, ids, 4, b) != 0 || + embed_model_encode_into(mprefixed, wprefixed, ids, 4, c) != 0) { fprintf(stderr, "embedding failed\n"); return 1; } float max_diff = 0.0f; + float prefix_diff = 0.0f; float norm = 0.0f; for (int i = 0; i < dim; i++) { - if (!isfinite(a[i]) || !isfinite(b[i])) { + if (!isfinite(a[i]) || !isfinite(b[i]) || !isfinite(c[i])) { fprintf(stderr, "non-finite embedding value\n"); return 1; } float d = fabsf(a[i] - b[i]); if (d > max_diff) max_diff = d; + d = fabsf(a[i] - c[i]); + if (d > prefix_diff) + prefix_diff = d; norm += b[i] * b[i]; } norm = sqrtf(norm); /* Embeddings are intentionally unnormalized; require a sane nonzero * magnitude and tight F32-vs-BF16 parity, not unit norm. */ - if (norm <= 1e-6f || max_diff > 2e-5f) { - fprintf(stderr, "bad parity: norm=%.9g max_diff=%.9g\n", norm, max_diff); + if (norm <= 1e-6f || max_diff > 2e-5f || prefix_diff != 0.0f) { + fprintf(stderr, "bad parity: norm=%.9g max_diff=%.9g prefix_diff=%.9g\n", norm, max_diff, + prefix_diff); return 1; } @@ -98,13 +110,17 @@ int main(void) { return 1; } - printf("ok: bf16 model parity dim=%d max_abs_diff=%.9g norm=%.9g\n", dim, max_diff, norm); + printf("ok: bf16 and model-prefix parity dim=%d max_abs_diff=%.9g norm=%.9g\n", dim, max_diff, + norm); embed_workspace_free(wf32); embed_workspace_free(wbf16); + embed_workspace_free(wprefixed); embed_model_free(mf32); embed_model_free(mbf16); + embed_model_free(mprefixed); free(a); free(b); + free(c); return 0; } diff --git a/tests/tiny_model.h b/tests/tiny_model.h index 42542c1..d2534d2 100644 --- a/tests/tiny_model.h +++ b/tests/tiny_model.h @@ -76,8 +76,8 @@ tm_write_config(const char *dir, const tm_dims_t *d, const char *model_type, int /* Write a safetensors file holding `specs` with tm_value() contents. * dtype: "F32" or "BF16". Returns 0 on success. */ -static int -tm_write_safetensors(const char *path, const char *dtype, const tm_spec_t *specs, int n_specs) { +static int tm_write_safetensors_with_prefix( + const char *path, const char *dtype, const tm_spec_t *specs, int n_specs, const char *prefix) { int bf16 = strcmp(dtype, "BF16") == 0; size_t esize = bf16 ? 2 : 4; @@ -89,15 +89,16 @@ tm_write_safetensors(const char *path, const char *dtype, const tm_spec_t *specs size_t n = (size_t)s->rows * (s->cols ? (size_t)s->cols : 1); size_t end = doff + n * esize; if (s->cols) - hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, - "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d,%d]," - "\"data_offsets\":[%zu,%zu]}", - t ? "," : "", s->name, dtype, s->rows, s->cols, doff, end); + hoff += + (size_t)snprintf(header + hoff, sizeof(header) - hoff, + "%s\"%s%s\":{\"dtype\":\"%s\",\"shape\":[%d,%d]," + "\"data_offsets\":[%zu,%zu]}", + t ? "," : "", prefix, s->name, dtype, s->rows, s->cols, doff, end); else hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, - "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d]," + "%s\"%s%s\":{\"dtype\":\"%s\",\"shape\":[%d]," "\"data_offsets\":[%zu,%zu]}", - t ? "," : "", s->name, dtype, s->rows, doff, end); + t ? "," : "", prefix, s->name, dtype, s->rows, doff, end); doff = end; } hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, "}"); @@ -129,9 +130,17 @@ tm_write_safetensors(const char *path, const char *dtype, const tm_spec_t *specs return 0; } +static int +tm_write_safetensors(const char *path, const char *dtype, const tm_spec_t *specs, int n_specs) { + return tm_write_safetensors_with_prefix(path, dtype, specs, n_specs, ""); +} + /* Write config.json + model.safetensors with the given dimensions into dir. * dtype: "F32" or "BF16". Returns 0 on success. */ -static int tm_write_model_dims(const char *dir, const char *dtype, const tm_dims_t *d) { +static int tm_write_model_dims_with_prefix(const char *dir, + const char *dtype, + const tm_dims_t *d, + const char *prefix) { int q = d->heads * d->head_dim, kv = d->kv_heads * d->head_dim; const tm_spec_t specs[TM_N_SPECS] = { {"embed_tokens.weight", d->vocab, d->hidden}, @@ -154,7 +163,16 @@ static int tm_write_model_dims(const char *dir, const char *dtype, const tm_dims return -1; snprintf(path, sizeof(path), "%s/model.safetensors", dir); - return tm_write_safetensors(path, dtype, specs, TM_N_SPECS); + return tm_write_safetensors_with_prefix(path, dtype, specs, TM_N_SPECS, prefix); +} + +static int tm_write_model_dims(const char *dir, const char *dtype, const tm_dims_t *d) { + return tm_write_model_dims_with_prefix(dir, dtype, d, ""); +} + +static inline int +tm_write_prefixed_model_dims(const char *dir, const char *dtype, const tm_dims_t *d) { + return tm_write_model_dims_with_prefix(dir, dtype, d, "model."); } static inline int tm_write_qwen3_model_dims(const char *dir, |
