diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-05-30 11:11:42 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-14 08:26:52 +0200 |
| commit | 44a235498ab54ffac97b7a0be659f6da83062d3f (patch) | |
| tree | 18dbe232204a443b6406ae4455996bffd576cdb4 /pplx_embed.h | |
| parent | 10bf114474d5b79193159e5c865734362cf7abfb (diff) | |
| download | ffwd-44a235498ab54ffac97b7a0be659f6da83062d3f.tar.gz ffwd-44a235498ab54ffac97b7a0be659f6da83062d3f.zip | |
dist: add distributed layer-slice foundation
Diffstat (limited to 'pplx_embed.h')
| -rw-r--r-- | pplx_embed.h | 31 |
1 files changed, 31 insertions, 0 deletions
diff --git a/pplx_embed.h b/pplx_embed.h index 0816e20..1a2479d 100644 --- a/pplx_embed.h +++ b/pplx_embed.h @@ -126,6 +126,37 @@ int pplx_model_embed_batch(const pplx_model_t *model, pplx_workspace_t *ws, float *out_embeddings); /* + * Run a contiguous transformer layer range for a packed/ragged batch. + * + * layer_start is inclusive and layer_end is exclusive. For a range beginning + * at layer 0, input_states must be NULL and every input must provide token + * ids. For a later range, input_states must contain the preceding range's + * packed [sum(n_tokens), hidden_size] output; token ids are not required. + * + * apply_final_norm is valid only for a range ending at the final transformer + * layer. out_states receives packed [sum(n_tokens), hidden_size] states. + * + * This is the core execution boundary used by layer-sharded inference. + */ +int pplx_model_forward_slice_batch(const pplx_model_t *model, + pplx_workspace_t *ws, + const pplx_input_t *inputs, int batch, + const float *input_states, + int layer_start, int layer_end, + int apply_final_norm, + float *out_states); + +/* + * Mean-pool packed final hidden states for a ragged batch. + * + * states is [sum(n_tokens), hidden_size]. seq_lengths contains one positive + * token count per sequence. The states must already include final RMSNorm. + */ +int pplx_pool_batch(const pplx_config_t *cfg, const float *states, + const int *seq_lengths, int batch, + float *out_embeddings); + +/* * Run the transformer forward pass WITHOUT pooling. * Returns the full [n_tokens * hidden_size] output after final RMSNorm. * Caller frees the returned buffer. NULL on error. |
