load embedding at current maxctx instead of max trained ctx by default

Author: LostRuins

klite.embd

@@ -3431,6 +3431,7 @@ Current version indicated by LITEVER below.
 
 		max_context_length: (localflag?4096:3072),
 		max_length: (localflag?512:256),
+		last_maxctx: 0,
 		auto_ctxlen: true,
 		auto_genamt: true,
 		rep_pen: 1.06,
@@ -11156,9 +11157,10 @@ Current version indicated by LITEVER below.
 											document.getElementById("max_length_slide").max = 4096;
 											document.getElementById("max_length_slide_label").innerText = 4096;
 										}
-										if(localflag && localsettings.max_context_length==defaultsettings.max_context_length && ep_maxctx>4096)
+										if(localflag && ep_maxctx>=4096 && localsettings.max_context_length<ep_maxctx && (localsettings.last_maxctx!=ep_maxctx || localsettings.max_context_length==defaultsettings.max_context_length))
 										{
 											localsettings.max_context_length = ep_maxctx;
+											localsettings.last_maxctx = ep_maxctx;
 										}
 									}).catch(error => {
 										console.log("Failed to get true max ctx: " + error);

koboldcpp.py

@@ -1901,7 +1901,7 @@ def embeddings_load_model(model_filename):
     inputs.flash_attention = False
     inputs.threads = args.threads
     inputs.use_mmap = args.usemmap
-    inputs.embeddingsmaxctx = args.embeddingsmaxctx
+    inputs.embeddingsmaxctx = (args.embeddingsmaxctx if args.embeddingsmaxctx else args.contextsize) # for us to clamp to contextsize if embeddingsmaxctx unspecified
     inputs = set_backend_props(inputs)
     ret = handle.embeddings_load_model(inputs)
     return ret