feat(bench): rocm bench load — GPU-saturating load generator for the dashboard by michaelroy-amd · Pull Request #90 · ROCm/rocm-cli

michaelroy-amd · 2026-07-08T21:21:24Z

Summary

Makes the bench section of the dashboard's Observe pane functional. The consumer half (CSV tailer → daemon → BenchRing → bench tab) already shipped; this adds the missing producer: rocm bench load, a minimal GPU-saturating load generator, then layers on server-side latency/peaks and a one-click TUI trigger.

The design was kept deliberately low-blast-radius: the generator appends normalized rows to the CSV a running daemon already tails, so the bench tab lights up live with no daemon/protocol/server/runner changes.

`rocm bench load` (Phase 1)

Fans out concurrent /chat/completions requests at a local OpenAI-compatible endpoint, measures client-side gen_tps/prompt_tps over a wall makespan, and writes one aggregate BenchmarkRow per concurrency cell.

tokio JoinSet + owned-permit Semaphore; per-request Result isolation (one refusal never zeroes a cell); makespan pinned before fan-out.
Append-only CSV: O_APPEND, one newline-terminated write_all per row (survives the live tailer), header-once, refuses to append on header mismatch (won't corrupt an external agent-bench CSV).
Top-level Command::Bench / bench load (the dash command is untouched). http://-only with a clear https:// rejection. Quality fields default to Unknown, so throughput rows never claim pass/fail.

Server-side latency + peaks (Phase 1.5, vLLM-only)

Scrapes the endpoint's /metrics before/after each cell for cumulative TTFT/TPOT histogram deltas, and runs a lightweight mid-cell poller (250 ms, 1 s scrape timeout) to capture true peak running/waiting requests.
CSV header extended 14→18 (max_running_reqs, max_waiting_reqs, ttft_ms, tpot_ms); shown in the bench detail panel. Non-vLLM endpoints degrade to None and never stall the sweep.

Auto-ramp + one-click TUI trigger (Phase 2)

--auto-ramp: doubling sweep 1→128 that stops on a pure should_stop_ramp (throughput plateau <5%, queue saturation guarded on running > 0, or hard cap); rows append incrementally so the tab fills live.
Press b on the Observe tab to open a bench-run form that launches rocm bench load via the existing job system, defaulting --out to the daemon-tailed CSV so results stream into the bench tab. Implemented with SpawnJob — no protocol change (an in-process Command::RunBench was evaluated and deliberately dropped for buying only ~1 s over the CSV tailer).

Dependencies

Zero new shipped deps. One new dev-dep: wiremock (HTTP stub for tests).

Test plan

cargo test -p rocm-dash-collectors -- --test-threads=1 — generator fields/tps, concurrency-cap enforcement, CSV round-trip + header guard, Prometheus delta + non-vLLM degradation, 18-col round-trip, deterministic should_stop_ramp (incl. an at-rest-peaks regression guard)
cargo test -p rocm-dash-tui -- --test-threads=1 — overlay invariants, SpawnJob args, https:// rejection, tailed-path default
cargo build -p rocm
cargo clippy --all-targets -- -D warnings — clean
Manual live smoke against a real rocm served model (#[ignore] in-tree)

Notes for reviewers

Full design + rationale (and the deferred items) live in plans/rocm-bench-load-generator.md.
TTFT/TPOT/peaks are vLLM-only by design (Prometheus /metrics); throughput works on any OpenAI-compatible endpoint.
The tool prints a "local smoke-test — not an official ROCm/AMD benchmark" caveat in --help and stdout.

EAI-7361 addendum — populate the dashboard Bench panel

Later commits on this branch fix the observability defect that left the dashboard's Bench panel permanently empty:

Root cause: bench_results_dir (the CSV file the daemon's CsvBenchTailer tails) defaulted to None in both the unified rocm-core config (DashboardDaemonConfig) and the standalone rocm-dash-core config (DaemonConfig), so the daemon tailed nothing unless a user hand-edited a config file — even after running rocm bench load. Additionally, rocm bench load with no --out wrote each run to a distinct timestamped file, so a plain CLI run was never on the tailed path.
Fix: both configs now default bench_results_dir to <data_dir>/bench/results.csv, resolved through the same AppPaths::discover() data-dir resolution the producer uses (so the ROCM_CLI_DATA_DIR override and host path normalization are honored identically on both sides). rocm bench load's no---out default now appends to that same file. A cross-binary regression test asserts producer default == daemon-tailed default resolve to the same absolute path.
Fixed the Bench empty-state copy (it referenced a nonexistent --bench-csv daemon flag; bench_results_dir is config-only) and reworded "directory" → "results file" since it is a single tailed file.

Intentional behavior change (reviewer note)

With a shared default output path (~/.rocm/bench/results.csv), a pre-existing CSV at that path whose header does not match the normalized benchmark schema now yields a loud HeaderMismatch refusal on the next rocm bench load, instead of the old behavior of silently writing a fresh timestamped file. This is intentional: a single canonical, daemon-tailed file is what makes the panel populate, and a schema mismatch should surface rather than fragment results across files. Users who want an ad-hoc separate file can still pass --out <path>.

Note: an earlier commit message on this branch stated the two configs' defaults were "identical / picked up here too" before they actually were — the extra /data/ segment in the rocm-dash-core path made them diverge. That mismatch is corrected by the alignment commit above; the defaults now genuinely resolve to the same path.

Rewrite the bench load-generator plan after an adversarial review pass against current main. Tightened to ~360 LOC / one new file: writer lives in collectors (daemon has no csv dep), top-level 'rocm bench load' verb (Dash untouched), one aggregate row per cell, distinct opt-in output CSV, http-only with a clear https rejection, and a fixed not-an-official- benchmark caveat in --help and stdout. Zero shipped deps; one dev-dep (wiremock). Supersedes the 2026-06-22 draft. Signed-off-by: Michael Roy <michael.roy@amd.com>

Phase 1 of the bench load generator: the consumer pipeline (CSV tailer -> daemon -> BenchRing -> bench tab) already shipped; this adds the missing producer. 'rocm bench load' fans out concurrent /chat/completions requests at a local OpenAI-compatible endpoint, measures client-side gen_tps and prompt_tps over a wall makespan, and appends one aggregate BenchmarkRow per concurrency cell to a distinct CSV a running daemon tails -- so the bench tab lights up live with zero daemon/protocol/UI changes. - collectors/bench_load.rs: LoadSpec, run_cell (JoinSet + owned-permit Semaphore, hand-serialized POST like lemonade.rs, per-request Result isolation, pinned makespan), run_and_append_csv (O_APPEND single newline-terminated write_all per row, header-if-empty, refuses on header mismatch). thiserror errors; no new shipped deps. - top-level Command::Bench/Load (Dash untouched); http-only with a clear https rejection; distinct opt-in --out (never the shared results dir); raw-throughput-vs-agent-workload caveat in --help and stdout. - tests T1-T5 (wiremock dev-dep) under --test-threads=1; T2 verifies the concurrency cap is actually enforced. Quality fields default to Unknown so throughput rows never claim pass/fail. Signed-off-by: Michael Roy <michael.roy@amd.com>

…gger (2) Phase 1.5 and Phase 2 of 'rocm bench load', making it feature-complete. Phase 1.5 (collectors, vLLM-only, zero new deps): each cell scrapes the endpoint's /metrics before/after for cumulative TTFT/TPOT histogram deltas, and a lightweight mid-cell poller (250ms, 1s scrape timeout, AtomicBool stop + join) tracks true peak running/waiting reqs. CSV header extended 14->18 (max_running_reqs, max_waiting_reqs, ttft_ms, tpot_ms); server-side gen_tps discarded in favour of the client-measured value. Non-vLLM endpoints degrade to None, never stalling the sweep. Latency shows in the bench detail panel. Phase 2: - auto-ramp (--auto-ramp): doubling sweep 1..128 that stops on a pure should_stop_ramp (plateau <5% gain, or queue saturation guarded on running>0, or hard cap); rows append incrementally so the tab fills live. - TUI trigger: press 'b' on Observe to open a bench-run form that spawns 'rocm bench load' via the existing job system, defaulting --out to the daemon-tailed CSV so rows stream into the bench tab. No protocol/server/ runner changes (Command::RunBench deliberately not built). Tests: Prometheus delta + non-vLLM degradation + 18-col round-trip + header guard; deterministic should_stop_ramp cases incl. an at-rest-peaks regression guard; overlay invariants + SpawnJob args + https rejection. All under --test-threads=1; clippy --all-targets -D warnings clean. Signed-off-by: Michael Roy <michael.roy@amd.com>

michaelroy-amd · 2026-07-08T22:08:10Z

CodeQL "uncontrolled data in path expression" — assessed and dismissed

All 18 rust/path-injection alerts on this PR have been reviewed and dismissed as not exploitable:

5 production hits (append_one_row / run_and_append_csv) — dismissed false positive. The flagged path is the operator's own --out flag (or a trusted config default) used to write the output CSV. It is invoked only from the interactive rocm bench load command, running with the user's own privileges — no trust boundary is crossed. The daemon/server never invoke this code path (Command::RunBench was intentionally not built), so there is no network- or IPC-reachable route. Writing the results file wherever the user chooses is the intended behavior of a benchmarking CLI; restricting it would be user-hostile with no security benefit.
13 test hits — dismissed used in tests. These are tempdir() paths constructed inside #[cfg(test)] unit tests, not reachable in production.

Each alert carries this rationale in its dismissal note. Happy to revisit if a future change routes an untrusted/remote value into the output path.

…pulates bench_results_dir (the file the daemon's CsvBenchTailer tails for normalized benchmark rows) defaulted to None in both the unified rocm-core config and the standalone rocm-dash config, so the Bench panel stayed empty even after running rocm bench load, unless a user hand-edited a config file to point at a CSV. Separately, rocm bench load with no --out wrote each run to a distinct self-labeled ~/.rocm/data/bench/rocm-bench-<timestamp>.csv file rather than the shared tailed path, so a plain CLI invocation was never visible to the daemon even when bench_results_dir was set. - Default bench_results_dir to <data_dir>/bench/results.csv in both DashboardDaemonConfig (rocm-core) and DaemonConfig (rocm-dash-core). - Default rocm bench load's --out (when omitted) to that same file so a plain CLI run appends to what the daemon already tails. - Fix the Bench tab's empty-state copy, which referenced a nonexistent --bench-csv daemon flag (bench_results_dir is config-only; there is no such CLI flag on any binary). Relates to EAI-7361 Signed-off-by: Michael Roy <michael.roy@amd.com>

…dir override Adversarial-review follow-up for EAI-7361: - Fix cross-binary path mismatch: rocm-dash-core's default resolved to $HOME/.rocm/data/bench/results.csv (extra /data/), but rocm-core's default_data_dir() is $HOME/.rocm itself, so the producer writes to $HOME/.rocm/bench/results.csv. Remove the stray /data/ segment so both configs resolve to the same absolute path the producer uses. - Honor ROCM_CLI_DATA_DIR on the daemon side: rocm-core's default now resolves through the same AppPaths::discover() the producer uses (not bare default_data_dir(), which ignores the override), and rocm-dash-core's default honors the override too. Previously, with the override set, a plain `rocm bench load` wrote a file the daemon never tailed. - Add a cross-binary regression test asserting the producer default (default_bench_csv_path) and the daemon-tailed default (DashboardDaemonConfig::bench_results_dir) resolve to the same path, plus an override-precedence test in rocm-dash-core. - Reword the Bench empty-state copy: "bench directory" -> "bench results file" (bench_results_dir is a single tailed file, not a directory). - Fix the --out help text's default path (drop the stray /data/). Relates to EAI-7361 Signed-off-by: Michael Roy <michael.roy@amd.com>

CI follow-up for EAI-7361: - Regenerate MANIFEST.md for the dependency additions on this branch (wiremock dev-dep tree: assert-json-diff, deadpool, num_cpus, etc.). Verified with `cargo xtask manifest --check`. - Harden CsvBenchTailer::drain against the CodeQL rust/path-injection alert at the File::open sink: the tailed path is derived from config / $ROCM_CLI_DATA_DIR, so route it through validated_read_path(), which canonicalizes the path (resolving symlinks and `..`) and rejects traversal before opening. The bench results file is intentionally user-configurable to any location, so canonicalization — not a fixed-root containment check — is the correct barrier. Canonicalization fails on an absent file exactly as the previous bare File::open did, so the daemon tick loop keeps tolerating a not-yet-created CSV. Relates to EAI-7361 Signed-off-by: Michael Roy <michael.roy@amd.com>

rominf

Solid, well-tested PR overall: clean copyright/SPDX headers, inline tests, wiremock correctly scoped as a dev-dep, and the CSV append/header-guard design is careful. The cross-binary default-path test is a nice touch. A few issues worth addressing before merge, ranked. Line numbers are on the PR head.

🔴 Correctness

1. --concurrency 0 hangs the process forever — crates/rocm-dash-collectors/src/bench_load.rs:126
Semaphore::new(concurrency as usize) with concurrency == 0 has zero permits, so every spawned task blocks on sem.acquire_owned().await (line 184) and no permit is ever released. js.join_next() never returns → run_cell hangs with no output and no timeout (the 5-min timeout is on the HTTP client, which is never reached). clap accepts 0 (and --concurrency 1,0,8) since there's no value_parser range. Reject non-positive concurrency at parse time, or skip/clamp 0 cells.

2. Unbounded "x".repeat(isl) allocation → OOM — bench_load.rs:188
"x".repeat(input_len as usize) is driven directly by the --isl u32 with no upper bound, allocated once per in-flight request. A fat-fingered --isl 2000000000 at concurrency 8 attempts ~16 GB simultaneously and aborts the process instead of returning a clean argument error. Bound --isl (and realistically --osl, --requests) to a sane maximum.

3. Auto-ramp saturation check compares independently-sampled peaks — bench_load.rs:443
should_stop_ramp stops when max_waiting_reqs >= max_running_reqs, but the poller (lines 150–163) tracks the two peaks as independent per-field maxima across the whole cell, not from one sample. A transient early queue spike (waiting peaks at 5 during dispatch) plus a later steady state (running peaks at 4) yields 5 >= 4 → premature stop, reporting saturation that never held at any single instant. Capture running+waiting from the same scrape to compare them.

4. A fully-failed cell permanently disables the plateau stop — bench_load.rs:474
In run_auto_ramp, if one cell has every request fail (n_success == 0 → gen_tps == None), prev_gen_tps = row.gen_tps sets it back to None; thereafter the plateau branch (Some(prev), Some(cur)) never matches, so the ramp keeps doubling concurrency up to 128, hammering an already-degraded endpoint. Only overwrite prev_gen_tps when the new value is Some.

🟠 Config / behavior change

5. Daemon now logs bench tailer drain failed every tick on a fresh install — crates/rocm-core/src/lib.rs (default_bench_results_dir) + crates/rocm-dash-collectors/src/bench_tail.rs
Flipping the default from None to Some(<data_dir>/bench/results.csv) means runner_options (apps/rocm/src/dash.rs:46) always builds a CsvBenchTailer, and the tick loop calls drain() every iteration (runner.rs:572). AppPaths::ensure() does not create bench/, so until the first rocm bench load the file is absent, canonicalize returns NotFound, and runner.rs:589 emits warn!("bench tailer drain failed") on every tick. Previously this was silent (no tailer). Suggest treating a not-yet-existing tailed file as an empty drain (return Ok(vec![]) on NotFound) rather than a warning.

6. The two default_bench_results_dir impls genuinely diverge despite the parity claim — crates/rocm-dash-core/src/config.rs:195
rocm-core's default goes through AppPaths::discover(), which — when ROCM_CLI_DATA_DIR is unset — picks up a config-driven managed root (configured_managed_root_from_config, e.g. setup.therock_venv) and the OS project-dirs fallback when $HOME is unset. rocm-dash-core's bench_results_dir_from_env only consults $ROCM_CLI_DATA_DIR/$HOME. So for a managed-root user (on Linux), rocm bench load writes <managed_root>/bench/results.csv but the standalone rocm-dash daemon tails $HOME/.rocm/bench/results.csv — the exact empty-panel bug this PR sets out to fix, reintroduced for the standalone daemon. The unified rocm binary path is consistent; the divergence is confined to standalone rocm-dash. Worth either sharing the resolution or documenting the limitation honestly (the current doc comment claims they "must resolve to the same absolute path").

🟡 Cleanup / altitude

7. Dead traversal guard — bench_tail.rs:76
The .components().any(|c| matches!(c, Component::ParentDir)) check runs after std::fs::canonicalize, which already resolves and removes all ... The rejection branch is unreachable, so the "path-injection barrier" the doc comment advertises never fires — the only real protection is that canonicalize resolves (not blocks) the path. Either drop the misleading check/comment or enforce containment under a known root.

8. Redundant header-mismatch pre-check — bench_load.rs:391
run_and_append_csv re-implements the header check (lines 391–403) that append_one_row already performs per row (349–363); run_auto_ramp omits it and works fine. ~13 duplicated lines that must stay byte-identical. Drop the outer copy.

9. Two reqwest::Clients rebuilt per cell — bench_load.rs:120
run_cell builds post_client + metrics_client on every call; auto-ramp calls it up to 8× → 16 client builds, discarding connection pools between cells. Build them once and pass them in.

Minor nits (non-blocking): metrics_url hand-rolls scheme/host parsing that rocm_dash_tui::llm / vllm_prom already do; the 250 ms poller 404-storms non-vLLM endpoints for the whole cell even though peaks stay None; and the field name bench_results_dir is a single file, forcing "despite the name" caveats in three places (a rename to bench_results_csv would delete all three).

📋 Repo hygiene — please remove `plans/`

10. plans/rocm-bench-load-generator.md (269 lines) should not be committed. It's self-described scratch/planning material (Owner: TBD, Supersedes the 2026-06-22 draft, "Phase 1 only") that goes stale immediately. The repo already .gitignores equivalent agent artifacts (active-implementation-log.md, current-user-asks.md, /.rocm-work/, /.claude), and AGENTS.md directs docs to README.md / docs/ / skills/ — plans/ is not a sanctioned location. There is no existing rule that explicitly forbids it, so this is a hygiene request plus a proposal:

Delete plans/ from this PR (move the design content into the PR description or docs/ if worth keeping).
Add plans/ to .gitignore.
Add a rule to AGENTS.md, e.g.:

Do not commit planning/scratch docs. Agent- or design-planning notes (e.g. plans/, implementation logs) must stay out of the repo — they go stale and bloat the tree. Durable design docs belong in docs/.

Validate load dimensions, keep queue saturation samples coherent, preserve ramp throughput across failed cells, and avoid warnings before the results CSV exists. Reuse HTTP clients across cells and remove transient planning material. Signed-off-by: Michael Roy <michael.roy@amd.com>

Keep the CodeQL path-flow barrier while treating a missing results file as an empty drain. Signed-off-by: Michael Roy <michael.roy@amd.com>

Signed-off-by: Michael Roy <michael.roy@amd.com>

…ator Signed-off-by: Michael Roy <michael.roy@amd.com>

volen-silo · 2026-07-14T09:21:31Z

🔴 Automated review · pr-review-watcher · 689a07c

Summary

Adds rocm bench load (a concurrency-sweep GPU load generator), vLLM-only server-side latency/peaks, a one-click TUI trigger, and (EAI-7361) defaults the daemon-tailed bench CSV so the dashboard's Bench panel populates. Verdict: Needs work — one blocking durable-config regression; the load generator, TUI wiring, CSV append/header-guard, and path-alignment logic are otherwise sound and well-tested. Verified: read all 18 changed files + collaborators (AppPaths/config serde, state StartJob, vllm_prom); confirmed producer (dash.rs::default_bench_csv_path) and unified-daemon (rocm-core::default_bench_results_dir) defaults both go through AppPaths::discover() so they resolve identically (incl. managed-root); confirmed the bit-packed peak poller, TTFT/TPOT delta math, and header-mismatch guard are correct; confirmed concurrency=0 is unreachable via the clap-validated CLI path. Did not run cargo build/cargo test (no toolchain in review env) — relied on source-level verification. Blocking: 1 · Non-blocking: 6.

🚫 Blocking (must fix before merge)

crates/rocm-core/src/lib.rs:3941 (default_bench_results_dir + the DashboardDaemonConfig.bench_results_dir field, ~3965) — a machine-specific absolute path is baked into the default and then persisted into config.json, silently reintroducing the exact bug EAI-7361 fixes. The field now defaults to Some(<data_dir>/bench/results.csv) and is serialized with skip_serializing_if = "Option::is_none" — which only skips None, so a Some(path) is always written. RocmCliConfig::save() (line 4133, to_vec_pretty(self)) is called from many mutation sites (therock, providers, automations, main). Flow: an existing pre-field config.json loads with the key absent → #[serde(default)] fills in the current machine's absolute path → the next save() pins that concrete path into config.json. Because serde default only fires on a missing key, the pinned path never re-resolves afterward. If $HOME/ROCM_CLI_DATA_DIR/the managed-root then differ (managed-root users, containers, config ported between machines), the daemon tails a stale path while rocm bench load writes elsewhere → Bench panel empty again, with no error. Fix: do not persist a derived per-machine path — resolve the default at read time in the consumer (e.g. in dash.rs::runner_options, fall back to default_bench_csv_path(paths) when the config value is None) and keep the serialized field None, mirroring how the standalone rocm-dash-core config resolves from env rather than storing an absolute path.

Non-blocking

apps/rocm/src/dash.rs:391 — explicit --out <path> does not create_dir_all the parent (the default-path branch does), so --out /new/dir/x.csv fails with an opaque os error 2; create the parent in both branches.
crates/rocm-dash-collectors/src/bench_load.rs:167 — run_cell/run_cell_with_clients hang forever on concurrency == 0 (Semaphore::new(0)); not reachable via the clap-validated CLI today, but the fn is pub — guard >= 1 and return BenchLoadError.
crates/rocm-dash-collectors/src/bench_load.rs:374 (append_one_row) — three separate opens (metadata → header read → append) are not atomic; two concurrent rocm bench load runs against the shared default path can both write the header (duplicate header parsed as a data row breaks CsvBenchTailer::drain). Open once and gate the header on file.metadata()?.len()==0; add an advisory lock if concurrent runs are supported.
crates/rocm-dash-tui/src/ui/bench_run.rs:206 — try_submit sets *br = None even when StartJob returns empty (a "bench-run" job already running is deduped in state.rs:197), so a re-submit silently closes the form with no job and no message; check fx.is_empty() and keep the form open with an error, like command_screen.rs.
crates/rocm-dash-tui/src/ui/bench_run.rs:69 — empty-Out hint claims the CLI default is ~/.rocm/bench/rocm-bench-<ts>.csv, but it is now the fixed <data_dir>/bench/results.csv; the hint is stale/misleading. Also BenchLoadError::HeaderMismatch gives no actionable "pass --out or delete the file" guidance for a user stuck on a pre-existing mismatched results.csv.
crates/rocm-dash-collectors/src/bench_load.rs:1087 — t11_auto_ramp_cap_reaches_128 asserts only non-empty + members-of-RAMP_SEQUENCE; it passes even if the ramp stops at concurrency=1, so it does not test its named claim. Also the cross-binary default-alignment tests silently skip when AppPaths::discover() fails rather than failing.

Signed-off-by: Michael Roy <michael.roy@amd.com>

…ator Signed-off-by: Michael Roy <michael.roy@amd.com>

michaelroy-amd · 2026-07-14T17:51:41Z

Blocking: derived bench_results_dir was serialized as a machine-specific absolute path.

Addressed in 8905e41: the config now keeps this field unset by default, and runner_options derives <data_dir>/bench/results.csv from the current AppPaths while preserving explicit overrides.

Also addressed the actionable non-blocking findings:

create parent directories for explicit and default --out paths;
reject run_cell(..., 0) at the public API boundary;
serialize cooperating CSV appenders with an advisory file lock so first writers emit one header;
keep the bench form open with an error when bench-run is already active;
correct the stale output hint and make header mismatch recovery actionable;
replace the nondeterministic auto-ramp cap test and remove skip-prone default-path tests.

The branch also includes the latest main at 53a56cc.

Validation:

cargo test --workspace --all-targets — 1478 passed, 6 ignored
cargo clippy --workspace --all-targets -- -D warnings — passed
python3 scripts/smoke_local.py — passed

Signed-off-by: Michael Roy <michael.roy@amd.com>

michaelroy-amd added 3 commits July 8, 2026 12:07

github-advanced-security AI found potential problems Jul 8, 2026

View reviewed changes

michaelroy-amd added 3 commits July 10, 2026 17:54

rominf requested changes Jul 13, 2026

View reviewed changes

github-advanced-security AI found potential problems Jul 13, 2026

View reviewed changes

Comment thread crates/rocm-dash-collectors/src/bench_tail.rs Fixed

michaelroy-amd added 2 commits July 13, 2026 16:52

fix(bench): preserve canonicalized tailer reads

88ce451

Keep the CodeQL path-flow barrier while treating a missing results file as an empty drain. Signed-off-by: Michael Roy <michael.roy@amd.com>

test(bench): avoid flagged temporary cleanup

f62593c

Signed-off-by: Michael Roy <michael.roy@amd.com>

michaelroy-amd enabled auto-merge July 14, 2026 02:37

Merge remote-tracking branch 'origin/main' into feat/bench-load-gener…

689a07c

…ator Signed-off-by: Michael Roy <michael.roy@amd.com>

michaelroy-amd added 2 commits July 14, 2026 10:26

fix(bench): address remaining review feedback

8905e41

Signed-off-by: Michael Roy <michael.roy@amd.com>

Merge remote-tracking branch 'origin/main' into feat/bench-load-gener…

53a56cc

…ator Signed-off-by: Michael Roy <michael.roy@amd.com>

github-advanced-security AI found potential problems Jul 14, 2026

View reviewed changes

test(bench): keep output path portable on Windows

a214725

Signed-off-by: Michael Roy <michael.roy@amd.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(bench): rocm bench load — GPU-saturating load generator for the dashboard#90

feat(bench): rocm bench load — GPU-saturating load generator for the dashboard#90
michaelroy-amd wants to merge 13 commits into
mainfrom
feat/bench-load-generator

michaelroy-amd commented Jul 8, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

michaelroy-amd commented Jul 8, 2026

Uh oh!

rominf left a comment

Uh oh!

Uh oh!

volen-silo commented Jul 14, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

michaelroy-amd commented Jul 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

michaelroy-amd commented Jul 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

rocm bench load (Phase 1)

Server-side latency + peaks (Phase 1.5, vLLM-only)

Auto-ramp + one-click TUI trigger (Phase 2)

Dependencies

Test plan

Notes for reviewers

EAI-7361 addendum — populate the dashboard Bench panel

Intentional behavior change (reviewer note)

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

michaelroy-amd commented Jul 8, 2026

CodeQL "uncontrolled data in path expression" — assessed and dismissed

Uh oh!

rominf left a comment

Choose a reason for hiding this comment

🔴 Correctness

🟠 Config / behavior change

🟡 Cleanup / altitude

📋 Repo hygiene — please remove plans/

Uh oh!

Uh oh!

volen-silo commented Jul 14, 2026

Summary

🚫 Blocking (must fix before merge)

Non-blocking

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

michaelroy-amd commented Jul 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

michaelroy-amd commented Jul 8, 2026 •

edited

Loading

`rocm bench load` (Phase 1)

📋 Repo hygiene — please remove `plans/`