llama : rename batch to ubatch (#9950)

This commit renames the member field batch in llm_build_context to
ubatch, and also the parameter batch in llama_build_graph, and
llama_set_inputs to ubatch.

The motivation for this change is to make the code more readable
(considering there are the structs llama_batch and llama_sbatch), and
consistent with other parts of the code base where parameters/fields of
type llama_ubatch are named ubatch.

diff --git a/src/llama.cpp b/src/llama.cpp
index 6a5c56a77e117ca2874cf57af1f2c9b1f12c0abe..7a5a46dce7dbc1ea99c48cd216b49e368311a7f9 100644 (file)
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -10017,7 +10017,7 @@ struct llm_build_context {
           llama_context  & lctx;
     const llama_hparams  & hparams;
     const llama_cparams  & cparams;
-    const llama_ubatch   & batch;
+    const llama_ubatch   & ubatch;
     const llama_kv_cache & kv_self;
 
     const int64_t n_embd;
@@ -10063,14 +10063,14 @@ struct llm_build_context {
     // TODO: consider making the entire interface noexcept
     llm_build_context(
         llama_context  & lctx,
-    const llama_ubatch & batch,
+    const llama_ubatch & ubatch,
     const llm_build_cb & cb,
                   bool   worst_case) :
         model            (lctx.model),
         lctx             (lctx),
         hparams          (model.hparams),
         cparams          (lctx.cparams),
-        batch            (batch),
+        ubatch           (ubatch),
         kv_self          (lctx.kv_self),
         n_embd           (hparams.n_embd),
         n_layer          (hparams.n_layer),
@@ -10092,7 +10092,7 @@ struct llm_build_context {
         beta_slow        (cparams.yarn_beta_slow),
         norm_eps         (hparams.f_norm_eps),
         norm_rms_eps     (hparams.f_norm_rms_eps),
-        n_tokens         (batch.n_tokens),
+        n_tokens         (ubatch.n_tokens),
         n_kv             (worst_case ? kv_self.size : kv_self.n),
         n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
         n_outputs_enc    (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd),
@@ -10461,7 +10461,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10621,7 +10621,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = model.type == MODEL_7B ? build_inp_pos() : nullptr;
@@ -10736,7 +10736,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10840,7 +10840,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10962,7 +10962,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // multiply by embedding_multiplier_scale of 78.38367176906169
         inpL = ggml_scale(ctx0, inpL, 78.38367176906169f);
@@ -11120,7 +11120,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -11242,7 +11242,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -11345,7 +11345,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11447,7 +11447,7 @@ struct llm_build_context {
         }
 
         // construct input embeddings (token, type, position)
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // token types are hardcoded to zero ("Sentence A")
         struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
@@ -11634,7 +11634,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11736,7 +11736,7 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11874,7 +11874,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12024,7 +12024,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12137,7 +12137,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12252,7 +12252,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12397,7 +12397,7 @@ struct llm_build_context {
         struct ggml_tensor * ffn_output;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12516,7 +12516,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12644,7 +12644,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12749,7 +12749,7 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12854,7 +12854,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12964,7 +12964,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13082,7 +13082,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13209,7 +13209,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // scale the input embeddings
         inpL = ggml_scale(ctx0, inpL, scale_embd);
@@ -13353,7 +13353,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // scale the input embeddings
         inpL = ggml_scale(ctx0, inpL, scale_embd);
@@ -13554,7 +13554,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
@@ -13662,7 +13662,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
@@ -13800,7 +13800,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13916,7 +13916,7 @@ struct llm_build_context {
         struct ggml_tensor * inpL;
 
         // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         struct ggml_tensor * state_copy = build_inp_s_copy();
         struct ggml_tensor * state_mask = build_inp_s_mask();
@@ -13928,7 +13928,7 @@ struct llm_build_context {
                     LLM_NORM_RMS, cb, il);
             cb(cur, "attn_norm", il);
 
-            cur = llm_build_mamba(ctx0, lctx, batch, gf, cur,
+            cur = llm_build_mamba(ctx0, lctx, ubatch, gf, cur,
                     state_copy, state_mask,
                     kv_head, n_kv, cb, il);
 
@@ -13974,7 +13974,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14131,7 +14131,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14259,7 +14259,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14378,7 +14378,7 @@ struct llm_build_context {
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14505,7 +14505,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14650,7 +14650,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14791,7 +14791,7 @@ struct llm_build_context {
         struct ggml_tensor * inpL;
 
         // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15006,7 +15006,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15160,7 +15160,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         GGML_ASSERT(lctx.is_encoding);
         struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
@@ -15292,7 +15292,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         GGML_ASSERT(!lctx.is_encoding);
         GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
@@ -15494,7 +15494,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -15586,7 +15586,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15700,7 +15700,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15824,7 +15824,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15944,11 +15944,11 @@ struct llm_build_context {
         // Token shift state dimensions should be 2 * n_emb
         GGML_ASSERT(n_embd == hparams.n_embd_k_s() / 2);
 
-        const int64_t n_seqs = batch.n_seqs;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_seqs = ubatch.n_seqs;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
         GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(batch.equal_seqs);
+        GGML_ASSERT(ubatch.equal_seqs);
         GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs);
 
         struct ggml_tensor * cur;
@@ -15956,7 +15956,7 @@ struct llm_build_context {
         struct ggml_tensor * state_copy = build_inp_s_copy();
         struct ggml_tensor * state_mask = build_inp_s_mask();
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
         inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
 
         for (int il = 0; il < n_layer; ++il) {
@@ -16070,7 +16070,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -16266,7 +16266,7 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {
 
 static struct ggml_cgraph * llama_build_graph(
          llama_context & lctx,
-    const llama_ubatch & batch,
+    const llama_ubatch & ubatch,
                   bool   worst_case) {
     const auto & model = lctx.model;
 
@@ -16288,7 +16288,7 @@ static struct ggml_cgraph * llama_build_graph(
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
         const bool full_offload = lctx.model.n_gpu_layers > (int)lctx.model.hparams.n_layer;
-        if (batch.n_tokens < 32 || full_offload) {
+        if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 for (auto * backend : lctx.backends) {
                     if (ggml_backend_supports_buft(backend, lctx.model.buft_layer[il].buft) &&
@@ -16303,7 +16303,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     struct ggml_cgraph * result = NULL;
 
-    struct llm_build_context llm(lctx, batch, cb, worst_case);
+    struct llm_build_context llm(lctx, ubatch, cb, worst_case);
 
     llm.init();
 
@@ -16554,7 +16554,7 @@ static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t
     return relative_bucket;
 }
 
-static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
+static void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) {
     //
     // set input data
     //
@@ -16563,28 +16563,28 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     const auto & cparams = lctx.cparams;
     const auto & kv_self = lctx.kv_self;
 
-    if (batch.token) {
-        const int64_t n_tokens = batch.n_tokens;
+    if (ubatch.token) {
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
+        ggml_backend_tensor_set(lctx.inp_tokens, ubatch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
     }
 
-    if (batch.embd) {
+    if (ubatch.embd) {
         const int64_t n_embd   = hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+        ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
     }
 
-    if (batch.pos && lctx.inp_pos) {
-        const int64_t n_tokens = batch.n_tokens;
+    if (ubatch.pos && lctx.inp_pos) {
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
     }
 
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_out_ids->buffer));
         int32_t * data = (int32_t *) lctx.inp_out_ids->data;
@@ -16593,10 +16593,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int i = 0; i < n_tokens; ++i) {
                 data[i] = i;
             }
-        } else if (batch.output) {
+        } else if (ubatch.output) {
             int32_t n_outputs = 0;
             for (int i = 0; i < n_tokens; ++i) {
-                if (batch.output[i]) {
+                if (ubatch.output[i]) {
                     data[n_outputs++] = i;
                 }
             }
@@ -16621,9 +16621,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache.
         if (cparams.causal_attn && !lctx.is_encoding) {
             const int64_t n_kv         = kv_self.n;
-            const int64_t n_tokens     = batch.n_tokens;
-            const int64_t n_seq_tokens = batch.n_seq_tokens;
-            const int64_t n_seqs       = batch.n_seqs;
+            const int64_t n_tokens     = ubatch.n_tokens;
+            const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+            const int64_t n_seqs       = ubatch.n_seqs;
 
 
             float * data     = nullptr;
@@ -16640,14 +16640,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             }
 
             // For causal attention, use only the previous KV cells
-            // of the correct sequence for each token of the batch.
+            // of the correct sequence for each token of the ubatch.
             // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
             for (int h = 0; h < 1; ++h) {
                 for (int s = 0; s < n_seqs; ++s) {
-                    const llama_seq_id seq_id = batch.seq_id[s][0];
+                    const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
                     for (int j = 0; j < n_seq_tokens; ++j) {
-                        const llama_pos pos = batch.pos[s*n_seq_tokens + j];
+                        const llama_pos pos = ubatch.pos[s*n_seq_tokens + j];
 
                         for (int i = 0; i < n_kv; ++i) {
                             float f;
@@ -16693,9 +16693,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
                 }
             }
         } else {
-            const int64_t n_tokens     = batch.n_tokens;
-            const int64_t n_seq_tokens = batch.n_seq_tokens;
-            const int64_t n_seqs       = batch.n_seqs;
+            const int64_t n_tokens     = ubatch.n_tokens;
+            const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+            const int64_t n_seqs       = ubatch.n_seqs;
             // when using kv cache, the mask needs to match the kv cache size
             const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens;
 
@@ -16705,7 +16705,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
 
             for (int h = 0; h < 1; ++h) {
                 for (int s1 = 0; s1 < n_seqs; ++s1) {
-                    const llama_seq_id seq_id = batch.seq_id[s1][0];
+                    const llama_seq_id seq_id = ubatch.seq_id[s1][0];
 
                     for (int j = 0; j < n_seq_tokens; ++j) {
                         const int32_t tj = s1*n_seq_tokens + j;
@@ -16715,10 +16715,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
-                                for (int s = 0; s < batch.n_seq_id[s0]; ++s) {
-                                    if (batch.seq_id[s0][s] == seq_id) {
+                                for (int s = 0; s < ubatch.n_seq_id[s0]; ++s) {
+                                    if (ubatch.seq_id[s0][s] == seq_id) {
                                         if (hparams.use_alibi) {
-                                            f = -std::abs(batch.pos[ti] - batch.pos[tj]);
+                                            f = -std::abs(ubatch.pos[ti] - ubatch.pos[tj]);
                                         } else {
                                             f = 0.0f;
                                         }
@@ -16740,9 +16740,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_mean);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
@@ -16753,12 +16753,12 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         std::vector<uint64_t> sum(n_tokens, 0);
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
 
-            sum[seq_id] += batch.n_seq_tokens;
+            sum[seq_id] += ubatch.n_seq_tokens;
         }
 
         std::vector<float> div(n_tokens, 0.0f);
@@ -16770,7 +16770,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         }
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
             for (int i = 0; i < n_seq_tokens; ++i) {
                 data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id];
@@ -16781,9 +16781,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     if (cparams.embeddings && (
                 cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
                 cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
@@ -16792,13 +16792,13 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
-                const llama_pos pos = batch.pos[s*n_seq_tokens + i];
+                const llama_pos pos = ubatch.pos[s*n_seq_tokens + i];
 
                 if (pos == 0) {
                     data[seq_id] = s*n_seq_tokens + i;
@@ -16808,9 +16808,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
@@ -16822,13 +16822,13 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         std::vector<int> last_row(n_tokens, -1);
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
-                const llama_pos pos = batch.pos[s*n_seq_tokens + i];
+                const llama_pos pos = ubatch.pos[s*n_seq_tokens + i];
 
                 if (pos >= last_pos[seq_id]) {
                     last_pos[seq_id] = pos;
@@ -16890,10 +16890,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (lctx.inp_pos_bucket) {
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer));
-        GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing
+        GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing
 
         int32_t * data = (int32_t *) lctx.inp_pos_bucket->data;
 
@@ -16902,7 +16902,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int h = 0; h < 1; ++h) {
                 for (int j = 0; j < n_tokens; ++j) {
                     for (int i = 0; i < n_kv; ++i) {
-                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
                     }
                 }
             }
@@ -16910,7 +16910,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int h = 0; h < 1; ++h) {
                 for (int j = 0; j < n_tokens; ++j) {
                     for (int i = 0; i < n_tokens; ++i) {
-                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(batch.pos[i], batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch.pos[i], ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
                     }
                 }
             }
@@ -16926,10 +16926,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
 
     if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) {
         const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer));
-        GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing
+        GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing
 
         float * data = (float *) lctx.inp_KQ_mask_cross->data;
 
@@ -16937,8 +16937,8 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int j = 0; j < n_tokens; ++j) {
                 for (int i = 0; i < n_output_enc; ++i) {
                     float f = -INFINITY;
-                    for (int s = 0; s < batch.n_seq_id[j]; ++s) {
-                        const llama_seq_id seq_id = batch.seq_id[j][s];
+                    for (int s = 0; s < ubatch.n_seq_id[j]; ++s) {
+                        const llama_seq_id seq_id = ubatch.seq_id[j][s];
                         if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) {
                             f = 0.0f;
                         }