lex-llm-vllm 0.3.11 → 0.3.13

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 66c6939650ded871419a3833d9d0a89e5c31a104f3acccbb065af86af17709cb
-  data.tar.gz: e639e39a60ff4c03838100bd361048997fecfe38e92fdd9f1763809958a43f8e
+  metadata.gz: a62d4fc4e5b1020ccf9a4bcb5ed9c03096923075c7111357648cc1b360dd804b
+  data.tar.gz: 1cb2fe7f84f9fcd44b0c04407e62310e5fa4fad59bdfd2bc65889bfa8b9bd213
 SHA512:
-  metadata.gz: cf346327446a5e3c7b8666c8ea04269d7fea32b7bbe89b9a0c11b7578fe3ca1a8f948cf4fc639265bfcfa6f0a65bca7007f24ce90146f6234f1589fcb6fe5c31
-  data.tar.gz: 1e324da7463bd1a63b14b5fde9b3f6030be0f8cdd933d802d350066beaf40fe8208998a298a848b00126bf5d1d68868b8384c2ed719fb863c825179aa8c92ced
+  metadata.gz: 0a68362c7660f1474b943d8d67948e71b3d2b23840b39835937839bf5d2542d0baaa05b569764f77e5268cde1848dfca20e2e16274de6f112756a8eaa1b79ccc
+  data.tar.gz: de8011c0d1468c8d3f35215f01b7ee3ffa182d47c56dc89ac5fd983dcbf6f9de52ef167fc04d7514b984423081b65f4a8a6f5b463dc6d4372bbcbcdd90d7aa75

data/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## [0.3.13] - 2026-07-14
+
+### Fixed
+- **Streaming dead stop on Qwen with thinking enabled.** `empty_delta?` only checked `delta['reasoning_content']`, missing vLLM's `delta['reasoning']` field. When the final chunk carried `finish_reason: "stop"` alongside a `reasoning` delta, the translator incorrectly treated it as an empty delta and fired the done chunk early — the content phase never started, producing an empty streaming response. Now checks both `reasoning_content` and `reasoning`. Also fixed `wire_metadata` to read from either field for sync responses.
+
+## [0.3.12] - 2026-07-05
+
+### Changed
+- Stop_reason mapping now uses the shared `Legion::Extensions::Llm::StopReasonMapping` mixin from lex-llm (>= 0.6.9) instead of a local `VLLM_STOP_REASON_MAP`. The local 3-entry map only recognized `tool_use` and silently fell back to `:end_turn` on the `tool_calls` finish_reason that OpenAI-compatible vLLM actually emits on tool turns — so vLLM tool completions were mis-reported as `end_turn`. The shared vocabulary maps both `tool_calls` and `tool_use` to `:tool_use` (plus `stop`/`end_turn`/`eos`, `length`/`max_tokens`, `stop_sequence`, `content_filter`), and is inherited by every provider so it no longer drifts per gem.
+
 ## [0.3.11] - 2026-06-20
 
 ### Fixed

data/lex-llm-vllm.gemspec

@@ -27,5 +27,5 @@ Gem::Specification.new do |spec|
   spec.add_dependency 'legion-logging', '>= 1.3.2'
   spec.add_dependency 'legion-settings', '>= 1.3.14'
   spec.add_dependency 'legion-transport', '>= 1.4.14'
-  spec.add_dependency 'lex-llm', '>= 0.6.0'
+  spec.add_dependency 'lex-llm', '>= 0.6.9'
 end

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

@@ -2,6 +2,7 @@
 
 require 'legion/extensions/llm/canonical'
 require 'legion/extensions/llm/responses/thinking_extractor'
+require 'legion/extensions/llm/stop_reason_mapping'
 require 'legion/json'
 require 'legion/logging'
 
@@ -25,13 +26,10 @@ module Legion
         # rubocop:disable Metrics/ClassLength, Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/PerceivedComplexity -- translator implementation
         class Translator
           include Legion::Logging::Helper
+          # Shared stop_reason vocabulary (tool_calls/tool_use/stop/length/...).
+          # vLLM is OpenAI-compatible and needs no additions; inherits the common map.
+          include Legion::Extensions::Llm::StopReasonMapping
 
-          # vLLM-specific stop_reason mapping (per conformance fixture stop_reason_matrix).
-          VLLM_STOP_REASON_MAP = {
-            'stop' => :end_turn,
-            'tool_use' => :tool_use,
-            'length' => :max_tokens
-          }.freeze
           FALLBACK_STOP_REASON = :end_turn
 
           # G18 parameter mapping: supported canonical params.
@@ -200,7 +198,7 @@ module Legion
               tool_calls_as_text: true,
               forced_tool_choice: true,
               thinking_tags: %w[think thinking],
-              stop_reason_map: VLLM_STOP_REASON_MAP,
+              stop_reason_map: stop_reason_map,
               streaming_token_usage: true
             }.freeze
           end
@@ -613,7 +611,8 @@ module Legion
 
           def wire_metadata(wire, message, _thinking_meta)
             meta = {}
-            meta[:reasoning_content] = message['reasoning_content'] if message['reasoning_content']
+            reasoning = message['reasoning_content'] || message['reasoning']
+            meta[:reasoning_content] = reasoning if reasoning
             raw_usage = wire['usage']
             if raw_usage.is_a?(Hash) && raw_usage['completion_tokens_details']
               meta[:completion_tokens_details] = raw_usage['completion_tokens_details']
@@ -657,7 +656,8 @@ module Legion
           def empty_delta?(delta)
             (delta['content'].nil? || delta['content'].to_s.empty?) &&
               (delta['tool_calls'].nil? || Array(delta['tool_calls']).empty?) &&
-              (delta['reasoning_content'].nil? || delta['reasoning_content'].to_s.empty?)
+              (delta['reasoning_content'].nil? || delta['reasoning_content'].to_s.empty?) &&
+              (delta['reasoning'].nil? || delta['reasoning'].to_s.empty?)
           end
 
           # Per-chunk think-tag extraction is structurally impossible while streaming:
@@ -694,7 +694,7 @@ module Legion
           def map_stop_reason(raw)
             return FALLBACK_STOP_REASON if raw.nil? || raw.to_s.empty?
 
-            VLLM_STOP_REASON_MAP.fetch(raw.to_s, FALLBACK_STOP_REASON)
+            stop_reason_lookup(raw) || FALLBACK_STOP_REASON
           end
 
           # ── JSON helpers ──

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

@@ -4,7 +4,7 @@ module Legion
   module Extensions
     module Llm
       module Vllm
-        VERSION = '0.3.11'
+        VERSION = '0.3.13'
       end
     end
   end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: lex-llm-vllm
 version: !ruby/object:Gem::Version
-  version: 0.3.11
+  version: 0.3.13
 platform: ruby
 authors:
 - LegionIO
@@ -71,14 +71,14 @@ dependencies:
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 0.6.0
+        version: 0.6.9
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 0.6.0
+        version: 0.6.9
 description: vLLM provider integration for the LegionIO LLM routing framework.
 email:
 - matthewdiverson@gmail.com