Parameterize inference-perf workload profiles

Replace hardcoded load stages and conversation_replay settings in the
agentic_code_generation and guide_predicted-latency-routing profiles with
${NUM_REQUESTS}, ${CONCURRENCY_LEVEL}/${CONCURRENCY}, and ${SEED}
environment variables so the profiles can be driven dynamically.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: Kaushik Mitra <kaushikmitra@google.com>

Author: kaushikmitr

workload/profiles/inference-perf/agentic_code_generation.yaml.in

@@ -3,16 +3,8 @@ load:
   num_workers: 8
   worker_max_concurrency: 1000
   stages:
-  - num_requests: 100
-    concurrency_level: 5
-  - num_requests: 200
-    concurrency_level: 10
-  - num_requests: 400
-    concurrency_level: 20
-  - num_requests: 600
-    concurrency_level: 30
-  - num_requests: 800
-    concurrency_level: 40
+  - num_requests: ${NUM_REQUESTS}
+    concurrency_level: ${CONCURRENCY_LEVEL}
 api:
   type: completion
   streaming: true
@@ -23,9 +15,9 @@ server:
 data:
   type: conversation_replay
   conversation_replay:
-    seed: 42
-    max_model_len: 262144 # Update based on the model's context length
-    num_conversations: 40
+    seed: ${SEED}
+    max_model_len: 262144
+    num_conversations: ${CONCURRENCY_LEVEL}
     shared_system_prompt_len: 3000
     dynamic_system_prompt_len:
       type: lognormal

workload/profiles/inference-perf/guide_predicted-latency-routing_1.yaml

@@ -3,8 +3,8 @@ load:
   num_workers: 6
   request_timeout: 300.0
   stages:
-    - num_requests: 192
-      concurrency_level: 32
+    - num_requests: ${NUM_REQUESTS}
+      concurrency_level: ${CONCURRENCY}
 api:
   type: completion
   streaming: true
@@ -18,7 +18,7 @@ tokenizer:
   data:
     type: conversation_replay
     conversation_replay:
-      seed: ${CONCURRENCY}         # unique seed per concurrency level
+      seed: ${SEED}         # unique seed per concurrency level
       num_conversations: ${CONCURRENCY}   # equals concurrency_level above
       shared_system_prompt_len: 3000
       dynamic_system_prompt_len: