Quelle  diff.patch   Sprache: unbekannt

diff --git a/cpp/actions.hpp b/cpp/actions.hpp
index ed79eff..079488c 100644
--- a/cpp/actions.hpp
+++ b/cpp/actions.hpp
@@ -173,6 +173,12 @@ glue_msg_load_res action_load(app_t &app, const char *req_raw)
     cparams.offload_kqv = req.offload_kqv.value;
   if (req.n_batch.not_null())
     cparams.n_batch = req.n_batch.value;
+  if (req.n_ubatch.not_null())
+    cparams.n_ubatch = req.n_ubatch.value;
+  if (req.flash_attn.not_null())
+    cparams.flash_attn = req.flash_attn.value;
+  if (req.n_threads_decoding.not_null())
+    cparams.n_threads = req.n_threads_decoding.value;
   if (req.n_seq_max.not_null())
     cparams.n_seq_max = req.n_seq_max.value;
   if (req.pooling_type.not_null())
diff --git a/cpp/glue.hpp b/cpp/glue.hpp
index 2cdc81b..e0f4106 100644
--- a/cpp/glue.hpp
+++ b/cpp/glue.hpp
@@ -513,6 +513,9 @@ struct glue_msg_load_req
   GLUE_FIELD_NULLABLE(int, yarn_orig_ctx)
   GLUE_FIELD_NULLABLE(str, cache_type_k)
   GLUE_FIELD_NULLABLE(str, cache_type_v)
+  GLUE_FIELD_NULLABLE(int, n_ubatch)
+  GLUE_FIELD_NULLABLE(bool, flash_attn)
+  GLUE_FIELD_NULLABLE(int, n_threads_decoding)
 };
 
 struct glue_msg_load_res
diff --git a/scripts/docker-compose.yml b/scripts/docker-compose.yml
index 27e8794..bab138a 100644
--- a/scripts/docker-compose.yml
+++ b/scripts/docker-compose.yml
@@ -26,7 +26,7 @@ services:
 
         # emcc --clear-cache
 
-        emcmake cmake ../..
+        emcmake cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_LLAMAFILE=OFF -DGGML_CPU_AARCH64=OFF ../..
         export EMCC_CFLAGS="$$SHARED_EMCC_CFLAGS"
         emmake make wllama -j
 
@@ -34,7 +34,7 @@ services:
         mkdir -p wasm/multi-thread
         cd wasm/multi-thread
         export EMCC_CFLAGS="" # temporary clear it
-        emcmake cmake ../..
+        emcmake cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_LLAMAFILE=OFF -DGGML_CPU_AARCH64=OFF ../..
         export EMCC_CFLAGS="$$SHARED_EMCC_CFLAGS -pthread -sUSE_PTHREADS=1 -sPTHREAD_POOL_SIZE=Module[\\\"pthreadPoolSize\\\"]"
         emmake make wllama -j
 
diff --git a/src/utils.ts b/src/utils.ts
index e0f217f..ff17a8c 100644
--- a/src/utils.ts
+++ b/src/utils.ts
@@ -114,7 +114,7 @@ export const sortFileByShard = (blobs: Blob[]): void => {
 export const delay = (ms: number) => new Promise((r) => setTimeout(r, ms));
 
 export const absoluteUrl = (relativePath: string) =>
-  new URL(relativePath, document.baseURI).href;
+  relativePath;
 
 export const padDigits = (number: number, digits: number) => {
   return (
diff --git a/src/wllama.ts b/src/wllama.ts
index c2996f5..1b6cb9e 100644
--- a/src/wllama.ts
+++ b/src/wllama.ts
@@ -565,6 +565,7 @@ export class Wllama {
       yarn_orig_ctx: config.yarn_orig_ctx,
       cache_type_k: config.cache_type_k as string,
       cache_type_v: config.cache_type_v as string,
+      ...config,
     });
     const loadedCtxInfo: LoadedContextInfo = {
       ...loadResult,
@@ -673,7 +674,7 @@ export class Wllama {
     await this.samplingInit(this.samplingConfig);
     const stopTokens = new Set(options.stopTokens ?? []);
     // process prompt
-    let tokens = await this.tokenize(prompt, true);
+    let tokens = Array.isArray(prompt) ? prompt : await this.tokenize(prompt, true);
     if (this.addBosToken && tokens[0] !== this.bosToken) {
       tokens.unshift(this.bosToken);
     }