aboutsummaryrefslogtreecommitdiff
path: root/embed.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-11 15:34:50 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:27:03 +0200
commit6b3c4bb1482917e4fb24480a1d082da5f191467b (patch)
tree6a29d58959d1c5d3a0febcc9cb06005ac5a009ed /embed.h
parent54e6e6bbef396eb1d38c09517d8157301fb8f151 (diff)
downloadffwd-6b3c4bb1482917e4fb24480a1d082da5f191467b.tar.gz
ffwd-6b3c4bb1482917e4fb24480a1d082da5f191467b.zip
dist: remove distributed layer-sharding (protocol, CLI flags, slice APIs)
Diffstat (limited to 'embed.h')
-rw-r--r--embed.h32
1 files changed, 0 insertions, 32 deletions
diff --git a/embed.h b/embed.h
index 2098e9e..37f58d5 100644
--- a/embed.h
+++ b/embed.h
@@ -88,17 +88,6 @@ typedef struct {
*/
PPLX_API pplx_model_t *pplx_model_load(const char *model_dir);
-/*
- * Load one contiguous transformer layer range for distributed execution.
- *
- * layer_start is inclusive and layer_end is exclusive. A range beginning at
- * zero also loads the token embedding table. A range ending at n_layers also
- * loads the final RMSNorm. Safetensors remain mmap-backed, so unselected
- * weights are not touched.
- */
-PPLX_API pplx_model_t *pplx_model_load_slice(const char *model_dir,
- int layer_start, int layer_end);
-
PPLX_API void pplx_model_free(pplx_model_t *model);
/*
@@ -158,27 +147,6 @@ PPLX_API int pplx_model_embed_batch(const pplx_model_t *model, pplx_workspace_t
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.
- */
-PPLX_API 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