diff options
| -rw-r--r-- | .gitignore | 4 | ||||
| -rw-r--r-- | main.c | 32 | ||||
| -rw-r--r-- | pplx_embed.c | 126 | ||||
| -rw-r--r-- | pplx_embed.h | 102 | ||||
| -rw-r--r-- | pplx_embed_mlx.c | 2 |
5 files changed, 88 insertions, 178 deletions
@@ -1,2 +1,6 @@ *.o pplx_embed +book/ +.venv/ +__pycache__/ +*.pyc @@ -53,7 +53,8 @@ static void print_usage(const char *prog) * ======================================================================== */ typedef struct { - pplx_ctx_t *ctx; + pplx_model_t *model; + pplx_workspace_t *workspace; #ifdef USE_MLX pplx_mlx_ctx_t *mlx_ctx; #endif @@ -74,7 +75,8 @@ static int engine_embed_batch(engine_t *e, const pplx_input_t *inputs, return pplx_mlx_embed_batch(e->mlx_ctx, inputs, batch, out_embeddings); else #endif - return pplx_embed_batch(e->ctx, inputs, batch, out_embeddings); + return pplx_model_embed_batch(e->model, e->workspace, + inputs, batch, out_embeddings); } static int tokenize_text(engine_t *e, const char *text, token_buf_t *out) @@ -457,14 +459,29 @@ int main(int argc, char *argv[]) #ifdef USE_MLX if (use_mlx) { e.mlx_ctx = pplx_mlx_load(model_dir); - if (!e.mlx_ctx) { fprintf(stderr, "failed to load model\n"); return 1; } + if (!e.mlx_ctx) { + fprintf(stderr, "failed to load model\n"); + qwen_tokenizer_free(tok); + return 1; + } e.dim = pplx_mlx_config(e.mlx_ctx)->hidden_size; } else #endif { - e.ctx = pplx_load(model_dir); - if (!e.ctx) { fprintf(stderr, "failed to load model\n"); return 1; } - e.dim = pplx_config(e.ctx)->hidden_size; + e.model = pplx_model_load(model_dir); + if (!e.model) { + fprintf(stderr, "failed to load model\n"); + qwen_tokenizer_free(tok); + return 1; + } + e.workspace = pplx_workspace_new(e.model); + if (!e.workspace) { + fprintf(stderr, "failed to allocate workspace\n"); + pplx_model_free(e.model); + qwen_tokenizer_free(tok); + return 1; + } + e.dim = pplx_model_config(e.model)->hidden_size; } if (verbose >= 1) fprintf(stderr, "Model: %d-dim, %.0f ms%s\n", @@ -478,7 +495,8 @@ int main(int argc, char *argv[]) rc = run_batch(&e, argc, argv, arg_start, print_embs, batch_size); /* Cleanup */ - if (e.ctx) pplx_free(e.ctx); + if (e.workspace) pplx_workspace_free(e.workspace); + if (e.model) pplx_model_free(e.model); #ifdef USE_MLX if (e.mlx_ctx) pplx_mlx_free(e.mlx_ctx); #endif diff --git a/pplx_embed.c b/pplx_embed.c index 7f48ba8..5fbc865 100644 --- a/pplx_embed.c +++ b/pplx_embed.c @@ -7,12 +7,52 @@ #include <string.h> #include <math.h> +typedef struct { + const float *wq; /* [q_dim, hidden] */ + const float *wk; /* [kv_dim, hidden] */ + const float *wv; /* [kv_dim, hidden] */ + const float *wo; /* [hidden, q_dim] */ + const float *q_norm; /* [head_dim] */ + const float *k_norm; /* [head_dim] */ + const float *input_norm; /* [hidden] */ + const float *post_attn_norm;/* [hidden] */ + const float *gate_proj; /* [intermediate, hidden] */ + const float *up_proj; /* [intermediate, hidden] */ + const float *down_proj; /* [hidden, intermediate] */ +} pplx_layer_t; + +typedef struct { + const float *embed_tokens; /* [vocab_size, hidden] */ + pplx_layer_t *layers; /* [n_layers] heap-allocated */ + const float *norm; /* [hidden] */ +} pplx_weights_t; + struct pplx_model { pplx_config_t config; pplx_weights_t weights; void *safetensors; /* multi_safetensors_t*, keeps mmap alive */ }; +struct pplx_workspace { + const pplx_model_t *model; /* immutable model this workspace belongs to */ + int buf_seq_cap; + float *x; /* [seq, hidden] */ + float *x_norm; /* [seq, hidden] */ + float *q; /* [seq, q_dim] */ + float *k; /* [seq, kv_dim] */ + float *v; /* [seq, kv_dim] */ + float *attn_out; /* [seq, q_dim] */ + float *proj_out; /* [seq, hidden] */ + float *ffn_gate; /* [seq, intermediate] */ + float *ffn_up; /* [seq, intermediate] */ + float *ffn_out; /* [seq, hidden] */ + + /* RoPE cosine/sine cache [n_pos, head_dim], grown lazily. */ + float *rope_cos; + float *rope_sin; + int rope_cache_cap; +}; + /* ======================================================================== * Globals * ======================================================================== */ @@ -70,7 +110,7 @@ static int parse_config(pplx_config_t *cfg, const char *model_dir) FILE *f = fopen(path, "rb"); if (!f) { - fprintf(stderr, "pplx_load: cannot open %s\n", path); + fprintf(stderr, "pplx_model_load: cannot open %s\n", path); return -1; } @@ -101,13 +141,13 @@ static int parse_config(pplx_config_t *cfg, const char *model_dir) /* Sanity checks */ if (cfg->hidden_size <= 0 || cfg->n_layers <= 0 || cfg->n_heads <= 0) { - fprintf(stderr, "pplx_load: invalid config in %s " + fprintf(stderr, "pplx_model_load: invalid config in %s " "(hidden=%d, layers=%d, heads=%d)\n", path, cfg->hidden_size, cfg->n_layers, cfg->n_heads); return -1; } if (cfg->n_layers > PPLX_MAX_LAYERS) { - fprintf(stderr, "pplx_load: too many layers (%d > %d)\n", + fprintf(stderr, "pplx_model_load: too many layers (%d > %d)\n", cfg->n_layers, PPLX_MAX_LAYERS); return -1; } @@ -240,14 +280,14 @@ pplx_model_t *pplx_model_load(const char *model_dir) /* Open safetensors (handles single file or multi-shard) */ multi_safetensors_t *ms = multi_safetensors_open(model_dir); if (!ms) { - fprintf(stderr, "pplx_load: failed to open safetensors in %s\n", model_dir); + fprintf(stderr, "pplx_model_load: failed to open safetensors in %s\n", model_dir); free(model); return NULL; } model->safetensors = ms; if (pplx_verbose >= 1) - fprintf(stderr, "pplx_load: loading weights from %s\n", model_dir); + fprintf(stderr, "pplx_model_load: loading weights from %s\n", model_dir); /* Token embeddings */ pplx_weights_t *w = &model->weights; @@ -292,7 +332,7 @@ pplx_model_t *pplx_model_load(const char *model_dir) if (!w->norm) goto fail; if (pplx_verbose >= 1) - fprintf(stderr, "pplx_load: %d layers loaded (%d-dim embeddings)\n", + fprintf(stderr, "pplx_model_load: %d layers loaded (%d-dim embeddings)\n", cfg->n_layers, cfg->hidden_size); return model; @@ -336,44 +376,6 @@ const pplx_config_t *pplx_model_config(const pplx_model_t *model) return model ? &model->config : NULL; } -const pplx_config_t *pplx_config(const pplx_ctx_t *ctx) -{ - return (ctx && ctx->model) ? pplx_model_config(ctx->model) : NULL; -} - -pplx_ctx_t *pplx_load(const char *model_dir) -{ - pplx_model_t *model = pplx_model_load(model_dir); - if (!model) return NULL; - - pplx_workspace_t *ws = pplx_workspace_new(model); - if (!ws) { - pplx_model_free(model); - return NULL; - } - - pplx_ctx_t *ctx = (pplx_ctx_t *)calloc(1, sizeof(pplx_ctx_t)); - if (!ctx) { - pplx_workspace_free(ws); - pplx_model_free(model); - return NULL; - } - - ctx->model = model; - ctx->workspace = ws; - ctx->config = model->config; - ctx->weights = model->weights; - return ctx; -} - -void pplx_free(pplx_ctx_t *ctx) -{ - if (!ctx) return; - pplx_workspace_free(ctx->workspace); - pplx_model_free(ctx->model); - free(ctx); -} - /* ======================================================================== * Packed forward pass * ======================================================================== */ @@ -624,42 +626,6 @@ float *pplx_model_embed(const pplx_model_t *model, pplx_workspace_t *ws, return emb; } -int pplx_forward_into(pplx_ctx_t *ctx, const int *token_ids, int n_tokens, - float *out_states) -{ - if (!ctx) return -1; - return pplx_model_forward_into(ctx->model, ctx->workspace, - token_ids, n_tokens, out_states); -} - -float *pplx_forward(pplx_ctx_t *ctx, const int *token_ids, int n_tokens) -{ - if (!ctx) return NULL; - return pplx_model_forward(ctx->model, ctx->workspace, token_ids, n_tokens); -} - -int pplx_embed_batch(pplx_ctx_t *ctx, const pplx_input_t *inputs, int batch, - float *out_embeddings) -{ - if (!ctx) return -1; - return pplx_model_embed_batch(ctx->model, ctx->workspace, - inputs, batch, out_embeddings); -} - -int pplx_embed_into(pplx_ctx_t *ctx, const int *token_ids, int n_tokens, - float *out_embedding) -{ - if (!ctx) return -1; - return pplx_model_embed_into(ctx->model, ctx->workspace, - token_ids, n_tokens, out_embedding); -} - -float *pplx_embed(pplx_ctx_t *ctx, const int *token_ids, int n_tokens) -{ - if (!ctx) return NULL; - return pplx_model_embed(ctx->model, ctx->workspace, token_ids, n_tokens); -} - /* ======================================================================== * Cosine similarity * ======================================================================== */ diff --git a/pplx_embed.h b/pplx_embed.h index 081b49c..6012576 100644 --- a/pplx_embed.h +++ b/pplx_embed.h @@ -37,67 +37,11 @@ typedef struct { } pplx_config_t; /* ======================================================================== - * Per-Layer Weights (all F32, direct mmap pointers - read-only) - * ======================================================================== */ - -typedef struct { - const float *wq; /* [q_dim, hidden] */ - const float *wk; /* [kv_dim, hidden] */ - const float *wv; /* [kv_dim, hidden] */ - const float *wo; /* [hidden, q_dim] */ - const float *q_norm; /* [head_dim] */ - const float *k_norm; /* [head_dim] */ - const float *input_norm; /* [hidden] */ - const float *post_attn_norm;/* [hidden] */ - const float *gate_proj; /* [intermediate, hidden] */ - const float *up_proj; /* [intermediate, hidden] */ - const float *down_proj; /* [hidden, intermediate] */ -} pplx_layer_t; - -/* ======================================================================== - * Full Model Weights - * ======================================================================== */ - -typedef struct { - const float *embed_tokens; /* [vocab_size, hidden] */ - pplx_layer_t *layers; /* [n_layers] heap-allocated */ - const float *norm; /* [hidden] */ -} pplx_weights_t; - -/* ======================================================================== * Model + workspace ownership * ======================================================================== */ typedef struct pplx_model pplx_model_t; - -typedef struct { - const pplx_model_t *model; /* immutable model this workspace belongs to */ - /* Working buffers, grown lazily to buf_seq_cap */ - int buf_seq_cap; - float *x; /* [seq, hidden] */ - float *x_norm; /* [seq, hidden] */ - float *q; /* [seq, q_dim] */ - float *k; /* [seq, kv_dim] */ - float *v; /* [seq, kv_dim] */ - float *attn_out; /* [seq, q_dim] */ - float *proj_out; /* [seq, hidden] */ - float *ffn_gate; /* [seq, intermediate] */ - float *ffn_up; /* [seq, intermediate] */ - float *ffn_out; /* [seq, hidden] */ - - /* RoPE cosine/sine cache [n_pos, head_dim], grown lazily */ - float *rope_cos; - float *rope_sin; - int rope_cache_cap; -} pplx_workspace_t; - -/* Convenience context: owns one model and one workspace. */ -typedef struct { - pplx_model_t *model; - pplx_workspace_t *workspace; - pplx_config_t config; /* backward-compatible snapshot */ - pplx_weights_t weights; /* backward-compatible read-only aliases */ -} pplx_ctx_t; +typedef struct pplx_workspace pplx_workspace_t; typedef struct { const int *ids; @@ -109,7 +53,7 @@ typedef struct { * ======================================================================== */ /* - * Load model from directory containing: + * Load immutable model data from a directory containing: * config.json - model hyperparameters * model.safetensors (or sharded) - F32 weights * vocab.json + merges.txt - BPE tokenizer @@ -117,30 +61,18 @@ typedef struct { * Reads config.json to determine model size (0.6B vs 4B). * Returns NULL on error. */ -pplx_ctx_t *pplx_load(const char *model_dir); - -/* Free all resources */ -void pplx_free(pplx_ctx_t *ctx); - -/* - * Load/free immutable model data only. A model owns config, mmap'd - * safetensors, and read-only weight pointers. It can be shared by multiple - * workspaces, but the model must outlive all workspaces created from it. - */ pplx_model_t *pplx_model_load(const char *model_dir); void pplx_model_free(pplx_model_t *model); /* - * Allocate/free mutable scratch buffers for a model. A workspace is not - * thread-safe, but multiple workspaces can use the same immutable model in - * different threads. + * Allocate/free mutable scratch buffers for a model. The model must outlive + * workspaces created from it. */ pplx_workspace_t *pplx_workspace_new(const pplx_model_t *model); void pplx_workspace_free(pplx_workspace_t *ws); -/* Config accessors for opaque-ish model/context ownership. */ +/* Config accessor for opaque model ownership. */ const pplx_config_t *pplx_model_config(const pplx_model_t *model); -const pplx_config_t *pplx_config(const pplx_ctx_t *ctx); /* * Compute embedding for a token sequence. @@ -153,14 +85,16 @@ const pplx_config_t *pplx_config(const pplx_ctx_t *ctx); * * Returns malloc'd float[hidden_size] (caller frees). NULL on error. */ -float *pplx_embed(pplx_ctx_t *ctx, const int *token_ids, int n_tokens); +float *pplx_model_embed(const pplx_model_t *model, pplx_workspace_t *ws, + const int *token_ids, int n_tokens); /* * Compute one embedding into caller-provided out[hidden_size]. * Returns 0 on success, -1 on error. */ -int pplx_embed_into(pplx_ctx_t *ctx, const int *token_ids, int n_tokens, - float *out_embedding); +int pplx_model_embed_into(const pplx_model_t *model, pplx_workspace_t *ws, + const int *token_ids, int n_tokens, + float *out_embedding); /* * Compute embeddings for a true packed/ragged batch. @@ -171,17 +105,9 @@ int pplx_embed_into(pplx_ctx_t *ctx, const int *token_ids, int n_tokens, * * Returns 0 on success, -1 on error. */ -int pplx_embed_batch(pplx_ctx_t *ctx, const pplx_input_t *inputs, int batch, - float *out_embeddings); - int pplx_model_embed_batch(const pplx_model_t *model, pplx_workspace_t *ws, const pplx_input_t *inputs, int batch, float *out_embeddings); -int pplx_model_embed_into(const pplx_model_t *model, pplx_workspace_t *ws, - const int *token_ids, int n_tokens, - float *out_embedding); -float *pplx_model_embed(const pplx_model_t *model, pplx_workspace_t *ws, - const int *token_ids, int n_tokens); /* * Run the transformer forward pass WITHOUT pooling. @@ -191,20 +117,16 @@ float *pplx_model_embed(const pplx_model_t *model, pplx_workspace_t *ws, * Useful for contextual (late-chunking) models where you need * per-token embeddings before splitting by separator positions. */ -float *pplx_forward(pplx_ctx_t *ctx, const int *token_ids, int n_tokens); +float *pplx_model_forward(const pplx_model_t *model, pplx_workspace_t *ws, + const int *token_ids, int n_tokens); /* * Run the transformer forward pass into caller-provided * out_states[n_tokens * hidden_size]. Returns 0 on success, -1 on error. */ -int pplx_forward_into(pplx_ctx_t *ctx, const int *token_ids, int n_tokens, - float *out_states); - int pplx_model_forward_into(const pplx_model_t *model, pplx_workspace_t *ws, const int *token_ids, int n_tokens, float *out_states); -float *pplx_model_forward(const pplx_model_t *model, pplx_workspace_t *ws, - const int *token_ids, int n_tokens); /* * Cosine similarity between two L2-normalized vectors. diff --git a/pplx_embed_mlx.c b/pplx_embed_mlx.c index d6fff8f..74eb51d 100644 --- a/pplx_embed_mlx.c +++ b/pplx_embed_mlx.c @@ -68,7 +68,7 @@ const pplx_config_t *pplx_mlx_config(const pplx_mlx_ctx_t *ctx) } /* ======================================================================== - * Load (reuses the same parse_config logic via pplx_load then copies) + * Load * ======================================================================== */ /* We duplicate a minimal config parser here so MLX can load standalone. |
