more consistency fixes

2026-09-05 20:41:18 +02:00 · 2025-08-13 19:28:53 +08:00
parent 955cf66bbc
commit 4b2ca1169c
3 changed files with 5 additions and 1 deletions
@@ -1898,7 +1898,7 @@ def embeddings_load_model(model_filename):
    inputs = embeddings_load_model_inputs()
    inputs.model_filename = model_filename.encode("UTF-8")
    inputs.gpulayers = (999 if args.embeddingsgpu else 0)
-    inputs.flash_attention = False
+    inputs.flash_attention = args.flashattention
    inputs.threads = args.threads
    inputs.use_mmap = args.usemmap
    inputs.embeddingsmaxctx = (args.embeddingsmaxctx if args.embeddingsmaxctx else args.contextsize) # for us to clamp to contextsize if embeddingsmaxctx unspecified