feat: validate ATOM and Infera on MI300X and MI355X

.github/AGENT_OPERATIONS.md

@@ -92,7 +92,7 @@ For multi-node configurations, `--all-evals` creates one eval job per engine top
 
 Single-node fixed-sequence results may include `power_valid`, `avg_power_w`, `avg_total_gpu_power_w`, `total_gpu_energy_j`, and joules per query/input/output/total token. Invalid telemetry records `power_valid: 0` without energy metrics and fails only with `REQUIRE_POWER=1`.
 
-Multinode disaggregated results add `prefill_gpu_energy_j`, `decode_gpu_energy_j`, `prefill_joules_per_input_token`, and `decode_joules_per_output_token`. Role energy covers the full formal benchmark window, not kernel-level phases.
+Multinode disaggregated results add `prefill_gpu_energy_j`, `decode_gpu_energy_j`, `prefill_avg_power_w`, `decode_avg_power_w`, `prefill_joules_per_input_token`, and `decode_joules_per_output_token`. Role energy covers the full formal benchmark window, not kernel-level phases, and the role watts are that energy divided by the same window and by the role's GPU count.
 
 For srt-slurm recipes, `telemetry: {provider: dcgm-power}` enables official energy collection. `runners/launch_gb200-nv.sh` and `runners/launch_gb300-nv.sh` are the source of truth for `POWER_SRT_SLURM_PIN`. CI derives `POWER_PRODUCER_SHA` from the launcher stamp. `utils/test_gb200_power_official_contract.py` and `utils/test_gb300_power_official_contract.py` enforce the recipe/launcher contract. Only `PRECISION=fp8` dcgm-power lanes are validated.
 

.github/workflows/e2e-tests.yml

@@ -261,6 +261,7 @@ jobs:
                   MULTI_AGENTIC_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' in x and x.get('run-eval', False)]))" | score_matrix multi-agentic-eval)
                   SINGLE=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))" | score_matrix single)
                   MULTI=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))" | score_matrix multi)
+                  EVALS=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))" | score_matrix eval)
                   MULTI_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))" | score_matrix multi-eval)
                   {
                     echo "agentic-config=$AGENTIC"

benchmarks/multi_node/srt-slurm-recipes/atom/qwen3-0.6b/mi300x/agg-2w-fixed-seq.yaml

@@ -0,0 +1,99 @@
+# Small two-worker aggregate correctness lane for the srt-slurm ATOM backend
+# and Infera's dynamic KV-aware router. This validates orchestration, not tuning.
+
+name: "mi300x-atom-qwen3-0.6b-agg-2w-fixed-seq"
+
+model:
+  path: "hf:Qwen/Qwen3-0.6B"
+  container: "infera-atom-v0.1.1"
+  precision: "fp16"
+
+identity:
+  model:
+    repo: "Qwen/Qwen3-0.6B"
+  container:
+    image: "rocm/infera:atom-v0.1.1"
+  frameworks:
+    atom: "0.1.4.dev113+g5837907f3"
+    infera: "0.0.0"
+
+slurm:
+  time_limit: "00:45:00"
+
+resources:
+  gpu_type: "mi300x"
+  gpus_per_node: 8
+  agg_nodes: 1
+  agg_workers: 2
+  gpus_per_agg: 1
+
+frontend:
+  type: infera
+  enable_multiple_frontends: false
+  env:
+    PYTHONPATH: "/atom-source:/infera-source"
+  args:
+    router-policy: kv-aware
+
+backend:
+  type: atom
+  enable_kv_events: true
+  aggregated_environment:
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    PYTHONPATH: "/atom-source:/infera-source"
+    PYTHONUNBUFFERED: "1"
+    OMP_NUM_THREADS: "1"
+  atom_config:
+    aggregated:
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.50
+      max-model-len: 2048
+      max-num-seqs: 8
+      block-size: 16
+      enforce-eager: true
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 240
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -euo pipefail
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    trap 'tar -C /logs -czf "'"${result_root}"'/runtime-logs.tar.gz" . 2>/dev/null || true' EXIT
+    for concurrency in 1 4; do
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+      --backend openai \
+      --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+      --endpoint /v1/completions \
+      --model Qwen/Qwen3-0.6B \
+      --tokenizer Qwen/Qwen3-0.6B \
+      --dataset-name random \
+      --random-input-len 128 \
+      --random-output-len 32 \
+      --random-prefix-len 96 \
+      --random-range-ratio 1.0 \
+      --random-num-workers 1 \
+      --num-warmups "${concurrency}" \
+      --num-prompts "$((concurrency * 4))" \
+      --max-concurrency "${concurrency}" \
+      --request-rate inf \
+      --ignore-eos \
+      --disable-tqdm \
+      --save-result \
+      --result-dir "${result_root}/fixed-seq" \
+      --result-filename "qwen3-0.6b-atom-agg-isl128-osl32-c${concurrency}.json";
+    done
+  env:
+    HF_HOME: /hf_hub_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    HUGGINGFACE_HUB_CACHE: /hf_hub_cache

benchmarks/multi_node/srt-slurm-recipes/atom/qwen3-0.6b/mi300x/disagg-1p1d-fixed-seq.yaml

@@ -0,0 +1,110 @@
+# Small two-node ATOM P/D correctness lane. Infera discovers both workers,
+# routes the completions API, and ATOM transfers KV through Mooncake.
+
+name: "mi300x-atom-qwen3-0.6b-disagg-1p1d-fixed-seq"
+
+model:
+  path: "hf:Qwen/Qwen3-0.6B"
+  container: "infera-atom-v0.1.1"
+  precision: "fp16"
+
+identity:
+  model:
+    repo: "Qwen/Qwen3-0.6B"
+  container:
+    image: "rocm/infera:atom-v0.1.1"
+  frameworks:
+    atom: "0.1.4.dev113+g5837907f3"
+    infera: "0.0.0"
+
+slurm:
+  time_limit: "00:45:00"
+
+resources:
+  gpu_type: "mi300x"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 1
+
+frontend:
+  type: infera
+  enable_multiple_frontends: false
+  env:
+    PYTHONPATH: "/atom-source:/infera-source"
+  args:
+    router-policy: kv-aware
+
+backend:
+  type: atom
+  connector: mooncake
+  # The stable atom-v0.1.1 image bundles Mooncake before ROCm DMA-BUF memory
+  # registration support. Use its supported TCP transport for this correctness
+  # lane; connection pooling is supplied by the srt-slurm ATOM adapter.
+  mooncake_protocol: tcp
+  enable_kv_events: true
+  prefill_environment: &worker_environment
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    PYTHONPATH: "/atom-source:/infera-source"
+    PYTHONUNBUFFERED: "1"
+    OMP_NUM_THREADS: "1"
+  decode_environment: *worker_environment
+  atom_config:
+    prefill: &worker_config
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.50
+      max-model-len: 2048
+      max-num-seqs: 8
+      block-size: 16
+      enforce-eager: true
+      no-enable_prefix_caching: true
+    decode: *worker_config
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 240
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -euo pipefail
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    trap 'tar -C /logs -czf "'"${result_root}"'/runtime-logs.tar.gz" . 2>/dev/null || true' EXIT
+    for concurrency in 1 4; do
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+      --backend openai \
+      --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+      --endpoint /v1/completions \
+      --model Qwen/Qwen3-0.6B \
+      --tokenizer Qwen/Qwen3-0.6B \
+      --dataset-name random \
+      --random-input-len 128 \
+      --random-output-len 32 \
+      --random-prefix-len 96 \
+      --random-range-ratio 1.0 \
+      --random-num-workers 1 \
+      --num-warmups "${concurrency}" \
+      --num-prompts "$((concurrency * 4))" \
+      --max-concurrency "${concurrency}" \
+      --request-rate inf \
+      --ignore-eos \
+      --disable-tqdm \
+      --save-result \
+      --result-dir "${result_root}/fixed-seq" \
+      --result-filename "qwen3-0.6b-atom-disagg-isl128-osl32-c${concurrency}.json";
+    done
+  env:
+    HF_HOME: /hf_hub_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    HUGGINGFACE_HUB_CACHE: /hf_hub_cache

benchmarks/multi_node/srt-slurm-recipes/atom/qwen3-0.6b/mi355x/agg-2w-fixed-seq.yaml

@@ -0,0 +1,99 @@
+# Small two-worker aggregate correctness lane for ATOM and Infera on MI355X.
+# It validates orchestration and KV-aware routing rather than performance.
+
+name: "mi355x-atom-qwen3-0.6b-agg-2w-fixed-seq"
+
+model:
+  path: "hf:Qwen/Qwen3-0.6B"
+  container: "infera-atom-v0.1.1"
+  precision: "fp16"
+
+identity:
+  model:
+    repo: "Qwen/Qwen3-0.6B"
+  container:
+    image: "rocm/infera:atom-v0.1.1"
+  frameworks:
+    atom: "0.1.4.dev113+g5837907f3"
+    infera: "0.0.0"
+
+slurm:
+  time_limit: "00:30:00"
+
+resources:
+  gpu_type: "mi355x"
+  gpus_per_node: 8
+  agg_nodes: 1
+  agg_workers: 2
+  gpus_per_agg: 1
+
+frontend:
+  type: infera
+  enable_multiple_frontends: false
+  env:
+    PYTHONPATH: "/atom-source:/infera-source"
+  args:
+    router-policy: kv-aware
+
+backend:
+  type: atom
+  enable_kv_events: true
+  aggregated_environment:
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache/hub"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache/hub"
+    PYTHONPATH: "/atom-source:/infera-source"
+    PYTHONUNBUFFERED: "1"
+    OMP_NUM_THREADS: "1"
+  atom_config:
+    aggregated:
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.50
+      max-model-len: 2048
+      max-num-seqs: 8
+      block-size: 16
+      enforce-eager: true
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 180
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -euo pipefail
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    trap 'tar -C /logs -czf "'"${result_root}"'/runtime-logs.tar.gz" . 2>/dev/null || true' EXIT
+    for concurrency in 1 4; do
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+      --backend openai \
+      --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+      --endpoint /v1/completions \
+      --model Qwen/Qwen3-0.6B \
+      --tokenizer Qwen/Qwen3-0.6B \
+      --dataset-name random \
+      --random-input-len 128 \
+      --random-output-len 32 \
+      --random-prefix-len 96 \
+      --random-range-ratio 1.0 \
+      --random-num-workers 1 \
+      --num-warmups "${concurrency}" \
+      --num-prompts "$((concurrency * 4))" \
+      --max-concurrency "${concurrency}" \
+      --request-rate inf \
+      --ignore-eos \
+      --disable-tqdm \
+      --save-result \
+      --result-dir "${result_root}/fixed-seq" \
+      --result-filename "qwen3-0.6b-atom-mi355x-agg-isl128-osl32-c${concurrency}.json";
+    done
+  env:
+    HF_HOME: /hf_hub_cache
+    HF_HUB_CACHE: /hf_hub_cache/hub
+    HUGGINGFACE_HUB_CACHE: /hf_hub_cache/hub