lex-llm-vllm 0.2.6 → 0.2.8

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 14dc7ee5334135f8eece7622a27cb57bad3c9c885cb6203de56006b5d2a1b353
-  data.tar.gz: 59ea84f7c50a9407da2af50c51f77ad26894d45b7f06d9c0223198a73a7e22a8
+  metadata.gz: 0ec67bb352cf0bd9dc3ae3f0102e7d26a27ebc8c64c56d98811742c4b75e4ae0
+  data.tar.gz: de092940c2c186e982b0d453b4f1f94157a868626e2a7b4b16cd6c0c1596b28c
 SHA512:
-  metadata.gz: 05c8fa912c908ec88943277dfd2e2b82f84bbd4ff6880d4e1f288f1397baf2cc1f72c2d20e894684be14be9f5f213e381ea5f524bfed3162e116de0271009f93
-  data.tar.gz: 14fffc18eff78d0c8751fbb96c9db6cbdd2773af66c46359b5934fad09d9a8fa37183a4f123f7cb75008e44cf821a21d97756d47633e44a50ab935f3f5743f0d
+  metadata.gz: aad15995696865b3cc0a6e42b6e627e2d4a5dbdb7768ec676d3867964fad107709ce10bfe908128a6575d8bc81d0a8ee9b97b45c554db093b12dbea5b67300c0
+  data.tar.gz: ecfe08d54a3113188c09f642e1b457af39eca29542315d529c477f3f2cac4096a7a1485fd2977fb6acdb7a4c6885f904510b9a3653eb9a966da5ffb6b60ac5ba

data/CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## 0.2.8 - 2026-05-07
+
+- Read vLLM thinking defaults from the active provider instance config so per-instance `enable_thinking` settings affect chat payloads.
+
+## 0.2.7 - 2026-05-07
+
+- Fix merge order in `discover_instances` so a user-supplied `tier:` in instance config is no longer clobbered by the `:direct` default.
+- Infer instance tier from endpoint URL in `normalize_instance_config`: `localhost`/`127.0.0.1`/`::1` → `:local`, any other host → `:direct`. Explicit `tier:` in config still wins.
+
 ## 0.2.6 - 2026-05-06
 
 - Load provider-owned fleet actors through the LegionIO subscription base and the canonical vLLM provider root.

data/lib/legion/extensions/llm/vllm/provider.rb

@@ -162,18 +162,25 @@ module Legion
           end
 
           def vllm_thinking_setting
+            instance_thinking_enabled? || global_thinking_enabled?
+          rescue StandardError => e
+            handle_exception(e, level: :debug, handled: true, operation: 'vllm.thinking_setting')
+            false
+          end
+
+          def instance_thinking_enabled?
+            return config.enable_thinking if config.respond_to?(:enable_thinking)
+
+            config.respond_to?(:[]) && config[:enable_thinking] == true
+          end
+
+          def global_thinking_enabled?
             return false unless defined?(Legion::Settings)
 
             vllm = Legion::Settings.dig(:llm, :providers, :vllm)
             return false unless vllm.is_a?(Hash)
 
-            vllm[:enable_thinking] == true ||
-              vllm['enable_thinking'] == true ||
-              vllm.dig(:instances, :default, :enable_thinking) == true ||
-              vllm.dig('instances', 'default', 'enable_thinking') == true
-          rescue StandardError => e
-            handle_exception(e, level: :debug, handled: true, operation: 'vllm.thinking_setting')
-            false
+            vllm[:enable_thinking] == true || vllm['enable_thinking'] == true
           end
 
           def parse_list_models_response(response, provider, capabilities)

data/lib/legion/extensions/llm/vllm/version.rb

@@ -4,7 +4,7 @@ module Legion
   module Extensions
     module Llm
       module Vllm
-        VERSION = '0.2.6'
+        VERSION = '0.2.8'
       end
     end
   end

data/lib/legion/extensions/llm/vllm.rb

@@ -14,6 +14,7 @@ module Legion
         extend Legion::Extensions::Llm::AutoRegistration
 
         PROVIDER_FAMILY = :vllm
+        DEFAULT_INSTANCE_TIER = { tier: :direct }.freeze
 
         def self.default_settings
           ::Legion::Extensions::Llm.provider_settings(
@@ -60,7 +61,7 @@ module Legion
           configured = CredentialSources.setting(:extensions, :llm, :vllm, :instances)
           if configured.is_a?(Hash)
             configured.each do |name, config|
-              instances[name.to_sym] = normalize_instance_config(config).merge(tier: :direct)
+              instances[name.to_sym] = DEFAULT_INSTANCE_TIER.merge(normalize_instance_config(config))
             end
           end
 
@@ -69,16 +70,31 @@ module Legion
 
         def self.normalize_instance_config(config)
           normalized = config.to_h.transform_keys(&:to_sym)
+          resolve_api_base_aliases(normalized)
+          normalized[:tier] ||= infer_tier_from_endpoint(normalized[:vllm_api_base])
+          normalized
+        end
+
+        def self.resolve_api_base_aliases(normalized)
           normalized[:vllm_api_base] ||= normalized.delete(:base_url)
           normalized[:vllm_api_base] ||= normalized.delete(:api_base)
           normalized[:vllm_api_base] ||= normalized.delete(:endpoint)
           normalized[:vllm_api_base] = normalize_api_base(normalized[:vllm_api_base]) if normalized[:vllm_api_base]
-          normalized
         end
 
         def self.normalize_api_base(url)
           url.to_s.sub(%r{/v1/?\z}, '')
         end
+
+        def self.infer_tier_from_endpoint(url)
+          return :direct if url.nil? || url.to_s.empty?
+
+          require 'uri'
+          host = URI.parse(url.to_s).host.to_s.downcase
+          %w[localhost 127.0.0.1 ::1].include?(host) ? :local : :direct
+        rescue URI::InvalidURIError
+          :direct
+        end
       end
     end
   end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: lex-llm-vllm
 version: !ruby/object:Gem::Version
-  version: 0.2.6
+  version: 0.2.8
 platform: ruby
 authors:
 - LegionIO