lex-llm-vllm 0.2.9 → 0.2.11

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c681beae79a3636380cbb8f75f3b0deb92722ee8dcfe150569944d4cd678ecd4
-  data.tar.gz: a072817f69752bde450cb67776b9b67021c680b5c744a9c05fddd6048821b871
+  metadata.gz: c4369ada04bb372dd59d3a57e2490c2b984c6ec1d7c4277c044b0f4126ba6136
+  data.tar.gz: 922947d49abecdbefbc7818dff56a82a35d1cae0978ebb94f5e9711f02a402c1
 SHA512:
-  metadata.gz: 9e4fdb96b3e7084371aa29f058072d2eb094c5872f0a10a4974c9bf2c16a6527d5e321563a7e24e0a9ee9ea52471ecf1c264a4e38e3c206930202739760d4135
-  data.tar.gz: f09e2f1c922a9466493281223371b57ff8b765850d15b9f783819551d9dc9aa1d8a36e5b40d6b9059d6dcf31998e2c6ca610a87c23c68ca70102e2e13fb5f193
+  metadata.gz: e2d9c3ffc63f2ba151573ebd898c0a08e8f73c52fe11ba38d452470e61569a33914d253ad8b1d08b235f90cfbc1f9613dea4bae5728d4b905539241e7034f9ff
+  data.tar.gz: 9cf36664ead33936f9c70ca3e892603ed6a66e1e3ef82560e7016e94207835ae6de4af7e7e9041e70eeb69b3117a48d780ca4a7e71aa3626ff8cc497800e8453

data/CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## 0.2.11 - 2026-05-21
+
+- Add `default_transport`/`default_tier` class declarations, remove duplicate instance methods
+- Add `model_allowed?` filtering in `discover_offerings`
+- Identity headers included via base provider
+- api_base reads from settings[:endpoint] fallback
+
+
+## 0.2.10 - 2026-05-13
+
+- Add `fetch_model_detail` to re-fetch `/v1/models` for `context_window` on a cache miss.
+- Pre-warm the model detail cache during offering discovery via `cache_set` using `model_detail_cache_key`.
+
 ## 0.2.9 - 2026-05-12
 
 - Route fleet actor load failures through `Legion::Logging::Helper` instead of direct warnings.

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

@@ -16,6 +16,8 @@ module Legion
           class << self
             def slug = 'vllm'
             def local? = false
+            def default_transport = :http
+            def default_tier = :direct
             def configuration_options = %i[vllm_api_base vllm_api_key]
             def configuration_requirements = []
             def capabilities = Capabilities
@@ -52,14 +54,14 @@ module Legion
           end
 
           def api_base
-            normalize_url(config.vllm_api_base || 'localhost:8000')
+            normalize_url(config.vllm_api_base || settings[:endpoint] || 'http://localhost:8000')
           end
 
           def headers
+            hdrs = identity_headers
             token = config.vllm_api_key
-            return {} if token.nil? || token.to_s.empty?
-
-            { 'Authorization' => "Bearer #{token}" }
+            hdrs['Authorization'] = "Bearer #{token}" unless token.nil? || token.to_s.empty?
+            hdrs
           end
 
           def health_url = '/health'
@@ -95,9 +97,13 @@ module Legion
                      else
                        Array(@cached_models)
                      end
-            models.map { |model_info| offering_from_model(model_info) }.tap do |offerings|
-              log.debug { "built #{offerings.size} vLLM offering(s) live=#{live}" }
+            offerings = models.filter_map do |model_info|
+              next unless model_allowed?(model_info.id)
+
+              offering_from_model(model_info)
             end
+            log.debug { "built #{offerings.size} vLLM offering(s) live=#{live}" }
+            offerings
           rescue StandardError => e
             handle_exception(e, level: :warn, handled: true, operation: 'vllm.discover_offerings')
             []
@@ -132,9 +138,28 @@ module Legion
             connection.post(with_query(wake_up_url, query), {}).body
           end
 
+          def fetch_model_detail(model_name)
+            # vLLM provides context_length via /v1/models during discovery.
+            # Re-fetch from the models endpoint if we need it outside discovery.
+            response = @connection.get(models_url)
+            models = response.body.fetch('data', [])
+            entry = models.find { |m| m['id'] == model_name.to_s }
+            return nil unless entry
+
+            ctx = entry['max_model_len']
+            ctx ? { context_window: ctx } : nil
+          rescue StandardError => e
+            handle_exception(e, level: :warn, handled: true, operation: 'vllm.fetch_model_detail',
+                                model: model_name)
+            nil
+          end
+
           private
 
           def offering_from_model(model_info)
+            ctx = model_info.context_length
+            cache_set(model_detail_cache_key(model_info.id), { context_window: ctx }, ttl: 86_400) if ctx
+
             Legion::Extensions::Llm::Routing::ModelOffering.new(
               provider_family: :vllm,
               instance_id: config.respond_to?(:instance_id) ? config.instance_id : :default,
@@ -143,19 +168,11 @@ module Legion
               model: model_info.id,
               usage_type: model_info.embedding? ? :embedding : :inference,
               capabilities: model_info.capabilities.map(&:to_s),
-              limits: { context_window: model_info.context_length }.compact,
-              metadata: { context_length: model_info.context_length }
+              limits: { context_window: ctx }.compact,
+              metadata: { context_length: ctx }
             )
           end
 
-          def offering_transport
-            config.respond_to?(:transport) ? config.transport : :http
-          end
-
-          def offering_tier
-            config.respond_to?(:tier) ? config.tier : :direct
-          end
-
           def render_payload(messages, tools:, temperature:, model:, stream:, schema:, thinking:, tool_prefs:) # rubocop:disable Metrics/ParameterLists
             payload = super
             payload.delete(:reasoning_effort)

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

@@ -4,7 +4,7 @@ module Legion
   module Extensions
     module Llm
       module Vllm
-        VERSION = '0.2.9'
+        VERSION = '0.2.11'
       end
     end
   end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: lex-llm-vllm
 version: !ruby/object:Gem::Version
-  version: 0.2.9
+  version: 0.2.11
 platform: ruby
 authors:
 - LegionIO