fix(benchmark): restore timeouts, add num_predict cap and deterministic options

Addresses quality=0 regression in the 20260309T174604 benchmark run:

- Restore benchmark_large_timeout to 480s (was inflated to 1200s as a
  workaround while hardware was degraded; hardware now restored to 26 tok/sec)
- Add benchmark_num_predict: 300 to cap output tokens, bounding worst-case
  completion time and preventing runaway generation on slow hardware
- Add options block to both fire-prompt tasks (node1 and node0) with
  num_predict, temperature: 0, and seed: 42 for deterministic, bounded results
- Fix latency calculation: subtract load_duration from total_duration so cold
  model loads do not inflate the TTFT measurement
- Restore benchmark timeout variables to pre-regression values (load: 180s,
  small: 90s, medium: 240s)

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

inventory/group_vars/all.yml

@@ -103,10 +103,11 @@ benchmark_skip_aliases:
 benchmark_small_max_gb: 10    # upper size boundary for small pass (< 10 GB), based on runtime RAM
 benchmark_medium_max_gb: 15   # upper size boundary for medium pass (10–15 GB), based on runtime RAM
 benchmark_size_overhead_factor: 1.2  # ollama list shows disk size; multiply by this to estimate runtime RAM
-benchmark_load_timeout: 300      # seconds — warm-up "Hi" prompt per model before benchmarking
-benchmark_small_timeout: 300     # seconds per request, small models
-benchmark_medium_timeout: 900    # seconds per request, medium models (15 min)
-benchmark_large_timeout: 1200    # seconds per request, large models (20 min)
+benchmark_load_timeout: 180      # seconds — warm-up "Hi" prompt per model before benchmarking
+benchmark_small_timeout: 90      # seconds per request, small models (<10 GB)
+benchmark_medium_timeout: 240    # seconds per request, medium models (10–15 GB)
+benchmark_large_timeout: 480     # seconds per request, large models (>15 GB)
+benchmark_num_predict: 300       # cap output tokens per prompt; bounds worst-case at 0.1 tok/sec to ~3000s, at 10 tok/sec to ~30s
 
 # Explicit category overrides applied before heuristics. Keys are model names as
 # returned by `ollama list`. Valid values: 'coding' or 'general'.

playbooks/03_benchmark.yml

@@ -265,7 +265,7 @@
           {%       set ns2.ttft_sum = ns2.ttft_sum + prompt_eval_duration %}
           {%       set ns2.ttft_count = ns2.ttft_count + 1 %}
           {%       if test_name == 'latency' %}
-          {%         set ns2.latency_ns = resp.total_duration | default(0) | int %}
+          {%         set ns2.latency_ns = ((resp.total_duration | default(0) | int) - (resp.load_duration | default(0) | int)) | abs %}
           {%       endif %}
           {%       set resp_len = response_text | length %}
           {%       if test_name in ['code_gen', 'debug', 'refactor'] %}

playbooks/_bench_tier_batch.yml

@@ -111,6 +111,10 @@
       model: "{{ item.0 }}"
       prompt: "{{ test_prompts[item.1].prompt }}"
       stream: false
+      options:
+        num_predict: "{{ benchmark_num_predict | int }}"
+        temperature: 0
+        seed: 42
     headers:
       Authorization: "Bearer {{ ollama_api_key }}"
     timeout: "{{ (benchmark_large_timeout | int) }}"
@@ -130,6 +134,10 @@
       model: "{{ item.0 }}"
       prompt: "{{ test_prompts[item.1].prompt }}"
       stream: false
+      options:
+        num_predict: "{{ benchmark_num_predict | int }}"
+        temperature: 0
+        seed: 42
     headers:
       Authorization: "Bearer {{ ollama_api_key }}"
     timeout: "{{ (benchmark_large_timeout | int) }}"