embedding_util 0.1.2 → 0.1.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: fb8276720e51a283fb4b8c70507ae54d25ea968caffe8142f636a6e6ea948b71
-  data.tar.gz: eccb6e6d11006238a1af1bfc25fb3b5146325bf78a1deb0f447a73541dec128d
+  metadata.gz: d7fe54bc241ead7c4f0aab2f252dfedff61627a1e6cbc3814e02096e455849bb
+  data.tar.gz: 540daea5a8b3ac13123b879f4644622bda81697981a4270867510be752bcb3c8
 SHA512:
-  metadata.gz: e625a2389828218a8f39cc4dc69d6ea965f59e032d190407668029df227f3c7ffee32ce30b64141d8a38654573ac591315c08696f051db3f68772efaf1b7bf34
-  data.tar.gz: 5da07a5487e7167128b67d8b6e39009d6a2fd4c0949daf1e61ae2cb9e98fa668beecd618713b1d0545daa47a86ac6f4e77f66d3429093c658852b594b4aa3cc5
+  metadata.gz: 8fd953d5fe00539c084aed32136ab17c3645a401ef768778baa2534b3125f8fcd57771839b11e1d1890896adbc60850bfb82e74c3b07708df04f9efff82eb5b5
+  data.tar.gz: 54d4812b7425a12ee64841e43604375267ef25ca1f6d0f9a661fd1e1e4d710fc8208904b9b836a233c10820be624ec248df063ab98cfb81709f8b6c8efd88693

data/CHANGELOG.md

@@ -1,3 +1,9 @@
+## [0.1.3] - 2026-06-10
+
+- Set self-hosted reranker `--batch-size` and `--ubatch-size` together
+- Retry managed reranker batch-size failures with both values raised to `4096`
+- Update endpoint guidance to recommend increasing both llama.cpp batch-size settings for app-managed rerankers
+
 ## [0.1.2] - 2026-06-10
 
 - Add self-hosted reranker recovery for llama.cpp physical batch-size failures

data/README.md

@@ -118,9 +118,9 @@ Reranker model:
 
 - repo: `ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF`
 - file: `qwen3-reranker-0.6b-q8_0.gguf`
-- server flags: `--reranking --ubatch-size 1024`
+- server flags: `--reranking --batch-size 1024 --ubatch-size 1024`
 
-For self-hosted rerankers, `embedding_util` starts with `--ubatch-size 1024`. If a larger rerank request exceeds llama.cpp's physical batch size, the managed reranker is restarted once with `config.reranker_max_ubatch_size`, which defaults to `4096`, and the request is retried.
+For self-hosted rerankers, `embedding_util` starts with `--batch-size 1024 --ubatch-size 1024`. If a larger rerank request exceeds llama.cpp's physical batch size, the managed reranker is restarted once with both values set to `config.reranker_max_ubatch_size`, which defaults to `4096`, and the request is retried.
 
 Do not combine embedding and reranking flags for this profile. Run separate local servers.
 

data/lib/embedding_util/profiles.rb

@@ -19,7 +19,7 @@ module EmbeddingUtil
         repo: "ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF",
         file: "qwen3-reranker-0.6b-q8_0.gguf",
         model: "qwen3-reranker-0.6b",
-        server_flags: ["--reranking", "--ubatch-size", "1024"]
+        server_flags: ["--reranking", "--batch-size", "1024", "--ubatch-size", "1024"]
       }
     )
 

data/lib/embedding_util/providers/endpoint.rb

@@ -153,8 +153,8 @@ module EmbeddingUtil
         message = "#{uri} returned #{response.code}: #{response.body}"
         return message unless reranker_batch_size_error?(path, response.body)
 
-        "#{message}. Restart the reranker server with a larger llama.cpp --ubatch-size; " \
-          "embedding_util-managed reranker servers use --ubatch-size 1024 by default."
+        "#{message}. Restart the reranker server with larger llama.cpp --batch-size and --ubatch-size values; " \
+          "embedding_util-managed reranker servers use 1024 by default and can retry with 4096."
       end
 
       def reranker_batch_size_error?(path, body)

data/lib/embedding_util/server_manager.rb

@@ -127,10 +127,10 @@ module EmbeddingUtil
       flags = server_model.settings.fetch(:server_flags)
       return flags unless server_model.capability == :reranker
 
-      with_ubatch_size(flags, config.reranker_ubatch_size)
+      with_reranker_batch_size(flags, config.reranker_ubatch_size)
     end
 
-    def with_ubatch_size(flags, size)
+    def with_reranker_batch_size(flags, size)
       filtered = []
       skip_next = false
       flags.each do |flag|
@@ -139,14 +139,14 @@ module EmbeddingUtil
           next
         end
 
-        if ["--ubatch-size", "-ub"].include?(flag)
+        if ["--batch-size", "-b", "--ubatch-size", "-ub"].include?(flag)
           skip_next = true
           next
         end
 
         filtered << flag
       end
-      filtered + ["--ubatch-size", size.to_s]
+      filtered + ["--batch-size", size.to_s, "--ubatch-size", size.to_s]
     end
 
     def required_port(host, port)

data/lib/embedding_util/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module EmbeddingUtil
-  VERSION = "0.1.2"
+  VERSION = "0.1.3"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: embedding_util
 version: !ruby/object:Gem::Version
-  version: 0.1.2
+  version: 0.1.3
 platform: ruby
 authors:
 - hmdne