[GLM-5.2] GB300 disagg iteration stats and AgentX worker metrics scraping / [GLM-5.2] GB300 disagg 迭代统计与 AgentX worker 指标采集

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/trtllm/gb300-fp4/agentx/disagg-variants.yaml

@@ -20,7 +20,7 @@ base:
     gpus_per_node: 4
   engine:
     type: trtllm
-    publish_events_and_metrics: false
+    publish_metrics: true
   roles:
     prefill:
       gpus: 4
@@ -71,6 +71,7 @@ base:
           event_buffer_max_size: 0
           free_gpu_memory_fraction: 0.75
           host_cache_size: 137438953472
+          iteration_stats_interval: 40
           tokens_per_block: 64
         max_batch_size: 256
         max_num_tokens: 8192
@@ -134,6 +135,7 @@ base:
           enable_block_reuse: false
           event_buffer_max_size: 0
           free_gpu_memory_fraction: 0.8
+          iteration_stats_interval: 40
           tokens_per_block: 64
         max_num_tokens: 128
         max_seq_len: 1048576
@@ -199,13 +201,15 @@ override_1p1d_tp8_c1_b1_mtp5:
   name: dynamo-disagg-gb300-1p1d-tp8-c1-b1-mtp5
   dynamo:
     source:
-      wheel: 1.4.0.dev20260807
+      git: https://github.com/cquil11/dynamo.git
+      rev: 2cbbdc863c336c2cfffa2a3668e2a6661801b259
   roles:
     prefill:
       nodes: 1
       workers: 1
       args:
         disable_overlap_scheduler: true
+        enable_iter_perf_stats: true
         scheduler_config:
           context_chunking_policy: EQUAL_PROGRESS
     decode:
@@ -217,6 +221,7 @@ override_1p1d_tp8_c1_b1_mtp5:
             - 1
         max_batch_size: 1
         moe_expert_parallel_size: 1
+        enable_iter_perf_stats: true
         sparse_attention_config:
           use_cute_dsl_topk: true
         tensor_parallel_size: 8

inferencex-e2e/perf-changelog.yaml

@@ -9211,3 +9211,14 @@
     - "Add --attention-config '{\"indexer_kv_dtype\":\"mxfp4\",\"indexer_sparse_logits\":true}' and --block-size 128 to enable the vllm-project/vllm#58671 ROCm paged MXFP4 sparse-logits indexer, replacing the dense fp8 indexer path. A live A/B test (TP2 c16, matched 900s window, vllm-project/vllm#58208 reverted via vllm-project/vllm#59125 so the dense fallback doesn't crash) measured +14.7/+14.9% p50/p90 interactivity and -9.5/-10.8% p50/p90 e2e latency over the dense path, with throughput/GPU unchanged."
     - "Drop c128 from both TP2 and TP4. Neither c128 point was on the Pareto frontier in #3555's run 36528242520: TP2 c64 dominated both (P90 E2EL 58 s against 185 s and 85 s, at 111k against 85k and 79k total tok/s/GPU)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3571
+
+- config-keys:
+    - glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Enable TensorRT-LLM iteration perf stats (enable_iter_perf_stats: true) on the prefill and decode workers of the GB300 1P1D TP8 concurrency 1 MTP5 variant (override_1p1d_tp8_c1_b1_mtp5), matching the other GLM-5.2 GB300 disagg variants. No serving, parallelism, precision or draft-model change."
+    - "Switch the c1 variant's Dynamo source from the 1.4.0.dev20260807 wheel to cquil11/dynamo@2cbbdc863c336c2cfffa2a3668e2a6661801b259, the build the other GB300 variants use."
+    - "Replace publish_events_and_metrics: false with publish_metrics: true in the shared engine base, so all six GB300 disagg variants publish TRT-LLM engine metrics (--publish-metrics) without KV-cache events."
+    - "Set kv_cache_config.iteration_stats_interval: 40 on the shared prefill and decode bases, so all six GB300 disagg variants inherit it."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3621