vllm-sr 0.1.0b2.dev20260204070724__py3-none-any.whl → 0.1.0b2.dev20260204090051__py3-none-any.whl

cli/merger.py

@@ -154,8 +154,20 @@ def translate_latency_signals(latencies: list) -> list:
     for signal in latencies:
         rule = {
             "name": signal.name,
-            "max_tpot": signal.max_tpot,
         }
+        # At least one of tpot_percentile or ttft_percentile should be set
+        if signal.tpot_percentile is not None and signal.tpot_percentile > 0:
+            rule["tpot_percentile"] = signal.tpot_percentile
+        if signal.ttft_percentile is not None and signal.ttft_percentile > 0:
+            rule["ttft_percentile"] = signal.ttft_percentile
+
+        # Validate that at least one is set
+        if "tpot_percentile" not in rule and "ttft_percentile" not in rule:
+            log.warn(
+                f"Latency signal '{signal.name}' has neither tpot_percentile nor ttft_percentile set, skipping"
+            )
+            continue
+
         if signal.description:
             rule["description"] = signal.description
         rules.append(rule)

cli/models.py

@@ -72,7 +72,8 @@ class Latency(BaseModel):
     """Latency signal configuration."""
 
     name: str
-    max_tpot: float
+    tpot_percentile: Optional[int] = None
+    ttft_percentile: Optional[int] = None
     description: str
 
 

cli/templates/config.template.yaml

@@ -177,14 +177,48 @@ signals:
     - name: "ja"
       description: "Japanese language queries"
 
-  # latency - Latency-based routing signals (TPOT-based)
+  # latency - Latency-based routing signals (TPOT and TTFT percentile-based)
+  # Percentile-based rules adapt to each model's actual performance distribution
+  # Works with any number of observations (1+): uses average for 1-2, percentile for 3+
+  # 
+  # ⚠️  RECOMMENDATION: Use BOTH tpot_percentile AND ttft_percentile for comprehensive latency evaluation
+  #    - TPOT: Measures token generation speed (throughput)
+  #    - TTFT: Measures first token latency (user-perceived latency)
+  #    - Together: Complete latency picture for optimal routing decisions
+  # 
+  # You CAN use only one if needed for specific use cases:
+  #   - TPOT only: Batch processing, cost optimization (throughput matters)
+  #   - TTFT only: Real-time chat (user perception matters)
+  #   - Both: RECOMMENDED for most applications (comprehensive evaluation)
+  # 
+  # At least one of tpot_percentile or ttft_percentile must be set (validation requirement)
+  # When both are set, model must meet BOTH thresholds (AND logic)
   latency:
-    - name: "low_latency"
-      max_tpot: 0.05  # 50ms per token
-      description: "For real-time chat applications requiring fast responses"
-    - name: "medium_latency"
-      max_tpot: 0.15  # 150ms per token
-      description: "For standard applications with moderate latency tolerance"
+    # Example 1: RECOMMENDED - Both TPOT and TTFT percentiles (comprehensive latency evaluation)
+    - name: "low_latency_comprehensive"
+      tpot_percentile: 10  # 10th percentile for TPOT (top 10% fastest token generation)
+      ttft_percentile: 10  # 10th percentile for TTFT (top 10% fastest first token)
+      description: "RECOMMENDED: For real-time applications - fast start and fast generation"
+    
+    # Example 2: Different percentiles for different priorities
+    - name: "balanced_latency"
+      tpot_percentile: 50  # Median TPOT (top 50%)
+      ttft_percentile: 10  # Top 10% TTFT (prioritize fast start)
+      description: "Prioritize fast start, accept moderate generation speed"
+    
+    # Example 3: TPOT percentile only (use case: batch processing, cost optimization)
+    # ⚠️  Note: Only using one metric is allowed but not recommended for most use cases
+    - name: "batch_processing_optimized"
+      tpot_percentile: 10  # 10th percentile for TPOT (top 10% fastest token generation)
+      # ttft_percentile: not set - acceptable for batch processing where throughput matters
+      description: "For batch processing where throughput (TPOT) is critical, TTFT less important"
+    
+    # Example 4: TTFT percentile only (use case: real-time chat where first token matters)
+    # ⚠️  Note: Only using one metric is allowed but not recommended for most use cases
+    - name: "chat_fast_start"
+      ttft_percentile: 10  # 10th percentile for TTFT (top 10% fastest first token)
+      # tpot_percentile: not set - acceptable for chat apps where user perception matters
+      description: "For chat applications where fast first token (TTFT) is critical for UX"
 
   # context - Context length signals (Token Count)
   context:
@@ -431,6 +465,39 @@ decisions:
           enabled: true
           similarity_threshold: 0.85
 
+  # Latency-based routing example: Route to models that meet latency requirements
+  - name: "low_latency_route"
+    description: "Route to models with low latency (fast TPOT and TTFT)"
+    priority: 90
+    rules:
+      operator: "AND"
+      conditions:
+        - type: "latency"
+          name: "low_latency_comprehensive"  # Requires both TPOT and TTFT percentiles
+    modelRefs:
+      - model: "openai/gpt-oss-120b"
+        use_reasoning: false
+    plugins:
+      - type: "system_prompt"
+        configuration:
+          system_prompt: "Provide fast, concise responses suitable for real-time applications."
+
+  - name: "fast_start_route"
+    description: "Route to models with fast first token (prioritize TTFT for chat apps)"
+    priority: 85
+    rules:
+      operator: "AND"
+      conditions:
+        - type: "latency"
+          name: "chat_fast_start"  # Only requires TTFT percentile
+    modelRefs:
+      - model: "openai/gpt-oss-120b"
+        use_reasoning: false
+    plugins:
+      - type: "system_prompt"
+        configuration:
+          system_prompt: "Start responding quickly. User is waiting for immediate feedback."
+
   # Size-aware routing example: Try smaller models first, escalate if confidence is low
   - name: "confidence_route"
     description: "Cost-efficient routing: try small model first, escalate if needed"

{vllm_sr-0.1.0b2.dev20260204070724.dist-info → vllm_sr-0.1.0b2.dev20260204090051.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vllm-sr
-Version: 0.1.0b2.dev20260204070724
+Version: 0.1.0b2.dev20260204090051
 Summary: vLLM Semantic Router - Intelligent routing for Mixture-of-Models
 Author: vLLM-SR Team
 License: Apache-2.0

{vllm_sr-0.1.0b2.dev20260204070724.dist-info → vllm_sr-0.1.0b2.dev20260204090051.dist-info}/RECORD

@@ -6,8 +6,8 @@ cli/defaults.py,sha256=x3mdZPBtI74bzgO5UDL-xoNV5jh9uPeSDcfgnkWHVyA,1969
 cli/docker_cli.py,sha256=kj3VvNfUJwcizZQj63mB2yLBBUbvZ9P3MtleIXhTp0M,25230
 cli/logo.py,sha256=I0qnCCGyOsHmN6MRgqa_c07MZSArv_6eHcHAL36V0Eg,1512
 cli/main.py,sha256=-PkXZw_z0VtzLbA5BCQlvLTYRuNbAT0CcBSFvjrsmBo,9137
-cli/merger.py,sha256=KMajL8tzdZeVRSdyq6MJt-DNhOQ5K_suMGx3jhl0odQ,16569
-cli/models.py,sha256=jrHiAqEIsWgu5O4_Y7T1xXLxFWlXuEtdr9RCLg1B0V0,13468
+cli/merger.py,sha256=Cd3DHt5_bMAw6oxSMtu5Debd6gCCRb0wjj-RTZSrZw4,17172
+cli/models.py,sha256=IOpG-XtLZtVDzQVIVd_PyLvMj_5BJDAt4R7kYBqr7Wo,13532
 cli/parser.py,sha256=lFPrROAYyc2cPoMMULV2sPid5IklIvc8zpDekj5Pocw,5653
 cli/utils.py,sha256=CHMwSZC-zxrxCnlZqiyzF2CToCVB5tHku2zyYig8fo0,3787
 cli/validator.py,sha256=G49I8_vyg0HvfI2LHwFlOagTwWAEsIkIDzX5riYoZhA,10651
@@ -19,7 +19,7 @@ cli/commands/serve.py,sha256=vsK3T4uWx49CPTgy4sc6oM7F5riq0sNbyiT7KbS37Wc,6104
 cli/commands/show_config.py,sha256=emkIdH9LKbzMzxna3Lxl-hEG85isfAXY5BlZjP_vaIM,4689
 cli/commands/show_defaults.py,sha256=r95KOHKQWeeNnoFO5EmCUZSPeOKb-5NzCoaZ6CaBcIw,701
 cli/commands/validate.py,sha256=Chm_MfyJMESiOo6IXiroYwMUkm8HCI26OHYYUBkGXGk,2773
-cli/templates/config.template.yaml,sha256=sS7ZJqrujBQgIs8E0vxUYcVofFL_SiGgW8XnNz5lrRE,18541
+cli/templates/config.template.yaml,sha256=8qnnyWiSLT7EkCFWK_O3qyHqvzfY_Q8GjmcbF13-e3k,22012
 cli/templates/envoy.template.yaml,sha256=j43mxQ3NcFMlmbGyrDM0-1rvQkC1N4MpXUC2o7SLT9g,8293
 cli/templates/generate_dashboard.py,sha256=gCui1rPrxUFIeioebyQ_JhL7zbjoEAg8GjUJcWXHcys,24233
 cli/templates/grafana-dashboard.serve.yaml,sha256=o1H5xx60esBzk19qnMQtX4LmL5DswpMvOO7cO09Q3kI,350
@@ -30,8 +30,8 @@ cli/templates/llm-router-dashboard.serve.json,sha256=pwnTjUh7z3_3LnIwtaLXjDWH4aH
 cli/templates/prometheus.serve.yaml,sha256=MGYq8dlRq_i2m5sogQ--kwTvJpkf44QQoCNoI7oyVT8,270
 cli/templates/router-defaults.yaml,sha256=a0riw9juttoBr4G7rtUEJA2FY9TgAvK-8mngWGCCIoc,9373
 cli/templates/tools_db.json,sha256=CPqPBkd5nc966m1YEozz06frrmv3Pd5rrkxKkO3rTiA,4537
-vllm_sr-0.1.0b2.dev20260204070724.dist-info/METADATA,sha256=Klc_-W-wKFC4HErcZy2jqrsSzfQQUSPI5JRB6MFWMCc,7298
-vllm_sr-0.1.0b2.dev20260204070724.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
-vllm_sr-0.1.0b2.dev20260204070724.dist-info/entry_points.txt,sha256=WhlBPbLHUpWUsMuUQX9cnvsYMf0ih5i57vvJ1jJNi0k,42
-vllm_sr-0.1.0b2.dev20260204070724.dist-info/top_level.txt,sha256=2ImG917oaVHlm0nP9oJE-Qrgs-fq_fGWgba2H1f8fpE,4
-vllm_sr-0.1.0b2.dev20260204070724.dist-info/RECORD,,
+vllm_sr-0.1.0b2.dev20260204090051.dist-info/METADATA,sha256=DoGzgyB-bYglhYLedrBqXLTlhB2RqF8rN1qosDC8SGQ,7298
+vllm_sr-0.1.0b2.dev20260204090051.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+vllm_sr-0.1.0b2.dev20260204090051.dist-info/entry_points.txt,sha256=WhlBPbLHUpWUsMuUQX9cnvsYMf0ih5i57vvJ1jJNi0k,42
+vllm_sr-0.1.0b2.dev20260204090051.dist-info/top_level.txt,sha256=2ImG917oaVHlm0nP9oJE-Qrgs-fq_fGWgba2H1f8fpE,4
+vllm_sr-0.1.0b2.dev20260204090051.dist-info/RECORD,,