Skip to content

feat: FMA actuation path classification, hit rates, and per-path timing #1429

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
aavarghese merged 5 commits into from
Jun 11, 2026
Merged

feat: FMA actuation path classification, hit rates, and per-path timing #1429

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
206e1aa
feat: Add FMA actuation path classification, hit rates, and per-path …
rubambiza May 29, 2026
d28086c
feat: Address PR review feedback on FMA metrics
rubambiza Jun 5, 2026
26aeab6
Merge branch 'main' into feature/fma-metrics-phases-1-2
rubambiza Jun 9, 2026
23beecc
feat: Replace maxSleepingInstances with maxInstances, add to metrics
rubambiza Jun 9, 2026
c56262e
refactor: Remove sleeperLimit from harness reporting pipeline
rubambiza Jun 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions config/templates/jinja/20_harness_pod.yaml.j2
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,6 +113,8 @@ spec:
value: "{{ fma.launcherConfigurator.port }}"
- name: LLMDBENCH_FMA_ITERATIONS
value: "{{ fma.iterations }}"
- name: LLMDBENCH_FMA_SLEEPER_LIMIT
Comment thread
aavarghese marked this conversation as resolved.
Outdated
value: "{{ fma.dualPod.sleeperLimit }}"
{% endif %}
Comment thread
aavarghese marked this conversation as resolved.
{% if harness.extraEnvVars is defined and harness.extraEnvVars %}
{% for env_entry in harness.extraEnvVars %}
Expand Down
187 changes: 142 additions & 45 deletions llmdbenchmark/analysis/benchmark_report/native_to_br0_1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,21 +52,29 @@ def _get_llmd_benchmark_envars() -> dict:
* int(os.environ.get("LLMDBENCH_VLLM_COMMON_REPLICAS", "-1")),
"accelerator": [
{
"model": os.environ.get("LLMDBENCH_VLLM_COMMON_AFFINITY", "").split(
":", 1
)[-1],
"model": os.environ.get(
"LLMDBENCH_VLLM_COMMON_AFFINITY", ""
).split(":", 1)[-1],
"count": int(
os.environ.get("LLMDBENCH_VLLM_COMMON_TENSOR_PARALLELISM", "-1")
os.environ.get(
"LLMDBENCH_VLLM_COMMON_TENSOR_PARALLELISM", "-1"
)
)
* int(os.environ.get("LLMDBENCH_VLLM_COMMON_DATA_PARALLELISM", "-1")),
* int(
os.environ.get(
"LLMDBENCH_VLLM_COMMON_DATA_PARALLELISM", "-1"
)
),
"parallelism": {
"tp": int(
os.environ.get(
"LLMDBENCH_VLLM_COMMON_TENSOR_PARALLELISM", "-1"
)
),
"dp": int(
os.environ.get("LLMDBENCH_VLLM_COMMON_DATA_PARALLELISM", "-1")
os.environ.get(
"LLMDBENCH_VLLM_COMMON_DATA_PARALLELISM", "-1"
)
),
},
}
Expand Down Expand Up @@ -97,14 +105,18 @@ def _get_llmd_benchmark_envars() -> dict:
},
},
"metadata": {
"load_format": os.environ.get("LLMDBENCH_VLLM_COMMON_VLLM_LOAD_FORMAT", ""),
"load_format": os.environ.get(
"LLMDBENCH_VLLM_COMMON_VLLM_LOAD_FORMAT", ""
),
"logging_level": os.environ.get(
"LLMDBENCH_VLLM_COMMON_VLLM_LOGGING_LEVEL", ""
),
"vllm_server_dev_mode": os.environ.get(
"LLMDBENCH_VLLM_COMMON_VLLM_SERVER_DEV_MODE", ""
),
"preprocess": os.environ.get("LLMDBENCH_VLLM_STANDALONE_PREPROCESS", ""),
"preprocess": os.environ.get(
"LLMDBENCH_VLLM_STANDALONE_PREPROCESS", ""
),
},
},
"metadata": {
Expand Down Expand Up @@ -150,89 +162,117 @@ def _get_llmd_benchmark_envars() -> dict:
"model": {"name": os.environ.get("LLMDBENCH_DEPLOY_CURRENT_MODEL", "")},
"host": {
"type": ["prefill"]
* int(os.environ.get("LLMDBENCH_VLLM_MODELSERVICE_PREFILL_REPLICAS", "-1"))
* int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_REPLICAS", "-1"
)
)
+ ["decode"]
* int(os.environ.get("LLMDBENCH_VLLM_MODELSERVICE_DECODE_REPLICAS", "-1")),
* int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_REPLICAS", "-1"
)
),
"accelerator": [
{
"model": os.environ.get("LLMDBENCH_VLLM_COMMON_AFFINITY", "").split(
":", 1
)[-1],
"model": os.environ.get(
"LLMDBENCH_VLLM_COMMON_AFFINITY", ""
).split(":", 1)[-1],
"count": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_TENSOR_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_TENSOR_PARALLELISM",
"-1",
)
)
* int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_DATA_LOCAL_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_DATA_LOCAL_PARALLELISM",
"-1",
)
),
"parallelism": {
"tp": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_TENSOR_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_TENSOR_PARALLELISM",
"-1",
)
),
"dp": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_DATA_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_DATA_PARALLELISM",
"-1",
)
),
"dpLocal": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_DATA_LOCAL_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_DATA_LOCAL_PARALLELISM",
"-1",
)
),
"workers": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_NUM_WORKERS_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_NUM_WORKERS_PARALLELISM",
"-1",
)
),
},
}
]
* int(os.environ.get("LLMDBENCH_VLLM_MODELSERVICE_PREFILL_REPLICAS", "-1"))
* int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_REPLICAS", "-1"
)
)
+ [
{
"model": os.environ.get("LLMDBENCH_VLLM_COMMON_AFFINITY", "").split(
":", 1
)[-1],
"model": os.environ.get(
"LLMDBENCH_VLLM_COMMON_AFFINITY", ""
).split(":", 1)[-1],
"count": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_TENSOR_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_TENSOR_PARALLELISM",
"-1",
)
)
* int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_DATA_LOCAL_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_DATA_LOCAL_PARALLELISM",
"-1",
)
),
"parallelism": {
"tp": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_TENSOR_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_TENSOR_PARALLELISM",
"-1",
)
),
"dp": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_DATA_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_DATA_PARALLELISM",
"-1",
)
),
"dpLocal": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_DATA_LOCAL_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_DATA_LOCAL_PARALLELISM",
"-1",
)
),
"workers": int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_NUM_WORKERS_PARALLELISM", "-1"
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_NUM_WORKERS_PARALLELISM",
"-1",
)
),
},
}
]
* int(os.environ.get("LLMDBENCH_VLLM_MODELSERVICE_DECODE_REPLICAS", "-1")),
* int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_REPLICAS", "-1"
)
),
},
"platform": {
"metadata": {
Expand All @@ -257,8 +297,16 @@ def _get_llmd_benchmark_envars() -> dict:
}
]
* (
int(os.environ.get("LLMDBENCH_VLLM_MODELSERVICE_PREFILL_REPLICAS", "-1"))
+ int(os.environ.get("LLMDBENCH_VLLM_MODELSERVICE_DECODE_REPLICAS", "-1"))
int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_PREFILL_REPLICAS", "-1"
)
)
+ int(
os.environ.get(
"LLMDBENCH_VLLM_MODELSERVICE_DECODE_REPLICAS", "-1"
)
)
),
},
},
Expand Down Expand Up @@ -1393,7 +1441,9 @@ def _stats(raw: dict | None, units: Units) -> dict | None:
"sessions_per_second": results.get("sessions_per_second"),
"session_duration": _stats(results.get("session_duration_sec"), Units.S),
"num_events": _stats(results.get("num_events"), Units.COUNT),
"num_events_cancelled": _stats(results.get("num_events_cancelled"), Units.COUNT),
"num_events_cancelled": _stats(
results.get("num_events_cancelled"), Units.COUNT
),
"total_input_tokens": _stats(results.get("total_input_tokens"), Units.COUNT),
"total_output_tokens": _stats(results.get("total_output_tokens"), Units.COUNT),
}
Expand All @@ -1418,7 +1468,8 @@ def _stats(raw: dict | None, units: Units) -> dict | None:
},
"requests": {
"total": results.get("total_events", 0),
"failures": results.get("total_events", 0) - results.get("total_events_completed", 0),
"failures": results.get("total_events", 0)
- results.get("total_events_completed", 0),
"input_length": {
"units": Units.COUNT,
"mean": get_nested(results, ["total_input_tokens", "mean"], 0),
Expand Down Expand Up @@ -2134,7 +2185,7 @@ def _import_categories(cat_list: list[dict[str, Any]]) -> list[dict[str, Any]]:
"args": engine["args"],
"metadata": {
"image": engine["image"],
}
},
}
engines.append(e)

Expand All @@ -2154,6 +2205,7 @@ def _import_categories(cat_list: list[dict[str, Any]]) -> list[dict[str, Any]]:
"deploy_methods": results["scenario"]["deploy_methods"],
"load_format": results["scenario"]["load_format"],
"sleep_mode": results["scenario"]["sleep_mode"],
"sleeper_limit": results["scenario"].get("sleeper_limit", 0),
Comment thread
aavarghese marked this conversation as resolved.
Outdated
"gpus": results["scenario"]["gpus"],
Comment thread
aavarghese marked this conversation as resolved.
},
},
Expand Down Expand Up @@ -2341,42 +2393,87 @@ def _import_categories(cat_list: list[dict[str, Any]]) -> list[dict[str, Any]]:
}
vllm_metadatas.append(metadata_dict)

results_dict["metrics"]["metadata"].append({"name": metrics_name, "value": vllm_metadatas})
results_dict["metrics"]["metadata"].append(
{"name": metrics_name, "value": vllm_metadatas}
)

metrics_name = "extra_metrics"
fma_metadatas = []
for extra_metric in results.get(metrics_name, []):
for extra_metric in results.get(metrics_name, []):
if extra_metric["name"] != "fma":
continue

metadata_dict = {"name": extra_metric["name"]}
iterations = []
for iteration in extra_metric.get("iterations", []):
it = { "iteration": { "units": Units.COUNT, "value": iteration["iteration"] } }
it = {"iteration": {"units": Units.COUNT, "value": iteration["iteration"]}}
launcher_infos = []
for launcher_info in iteration.get("launcher_infos", []):
info = { "name": launcher_info["name"] }
info = {"name": launcher_info["name"]}

requester_info = launcher_info["requester_info"]
ri = { "name": requester_info["name"] }
ri["creation_timestamp"] = { "units": Units.S, "value": requester_info["creation_timestamp"]}
ri["ready_timestamp"] = { "units": Units.S, "value": requester_info["ready_timestamp"]}
ri["dual_label_timestamp"] = { "units": Units.S, "value": requester_info["dual_label_timestamp"]}
ri = {"name": requester_info["name"]}
ri["creation_timestamp"] = {
"units": Units.S,
"value": requester_info["creation_timestamp"],
}
ri["ready_timestamp"] = {
"units": Units.S,
"value": requester_info["ready_timestamp"],
}
ri["dual_label_timestamp"] = {
"units": Units.S,
"value": requester_info["dual_label_timestamp"],
}
info["requester_info"] = ri

info["actuation_condition"] = launcher_info["actuation_condition"]
info["launcher_endpoint"] = launcher_info["launcher_endpoint"]
info["vllm_endpoint"] = launcher_info["vllm_endpoint"]
info["ttft"] = { "units": Units.S, "value": launcher_info["ttft"]}
info["ttft"] = {"units": Units.S, "value": launcher_info["ttft"]}
info["launcher_creation_timestamp"] = {
"units": Units.S,
"value": launcher_info.get("launcher_creation_timestamp", 0.0),
}
info["launcher_node"] = launcher_info.get("launcher_node", "")
if launcher_info.get("t_wake") is not None:
info["t_wake"] = {
"units": Units.S,
"value": launcher_info["t_wake"],
}
if launcher_info.get("t_instance_create") is not None:
info["t_instance_create"] = {
"units": Units.S,
"value": launcher_info["t_instance_create"],
}
if launcher_info.get("t_cold_launcher") is not None:
info["t_cold_launcher"] = {
"units": Units.S,
"value": launcher_info["t_cold_launcher"],
}
launcher_infos.append(info)

it["launcher_infos"] = launcher_infos
it["hot_hit_rate"] = {
"units": Units.COUNT,
"value": iteration.get("hot_hit_rate", 0.0),
}
it["warm_hit_rate"] = {
"units": Units.COUNT,
"value": iteration.get("warm_hit_rate", 0.0),
}
it["cold_launcher_hit_rate"] = {
"units": Units.COUNT,
"value": iteration.get("cold_launcher_hit_rate", 0.0),
}
iterations.append(it)

metadata_dict["iterations"] = iterations
fma_metadatas.append(metadata_dict)

results_dict["metrics"]["metadata"].append({"name": metrics_name, "value": fma_metadatas})
results_dict["metrics"]["metadata"].append(
{"name": metrics_name, "value": fma_metadatas}
)

update_dict(br_dict, results_dict)

Expand Down
Loading