Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Stop the stall-window comparison being decided by one noisy sample (#10246)
Daniel Han2ed0da03e34e
diff --git a/studio/frontend/tests/background-load-notice.test.ts b/studio/frontend/tests/background-load-notice.test.tsindex a31bd1a0c10..c655f132d8c 100644--- a/studio/frontend/tests/background-load-notice.test.ts+++ b/studio/frontend/tests/background-load-notice.test.ts@@ -361,18 +361,39 @@ test("a healthy read resets the stall window", async () => {// So measure the reset against the same loop without one. The comparison// divides the platform out: whatever a poll costs, a window restarted halfway// through must carry the loop about half as far again.- const withoutReset = await readsBeforeSettling(0);- assert.ok(- withoutReset > 8,- `the baseline is too short to compare against: ${withoutReset} reads. ` +- "Raise STALL_MS.",- );+ // The two halves of the comparison are separate wall-clock runs, so anything+ // that slows the runner between them skews the ratio. A windows-latest runner+ // is shared and noisy enough to do that: this failed on main at 56 reads with+ // a reset against 47 without, a ratio of 1.19 where a half-way restart should+ // give about 1.5. One sample decided it, and the sample was unlucky.+ //+ // Take the majority of three rounds rather than one sample, and NOT the best+ // of three. The noise runs in both directions, so with the reset removed --+ // where the two loops are the same loop and should score about 1.0 -- a round+ // that happened to schedule the second run faster could clear the bar on its+ // own, and accepting the most favourable round would make three chances at+ // that instead of one. A majority is strictly stronger than the single sample+ // this replaces in both directions: it takes two unlucky rounds to fail a+ // healthy reset, and two lucky ones to pass a missing one.+ const rounds: string[] = [];+ let cleared = 0;+ for (let round = 1; round <= 3; round += 1) {+ const withoutReset = await readsBeforeSettling(0);+ assert.ok(+ withoutReset > 8,+ `the baseline is too short to compare against: ${withoutReset} reads. ` ++ "Raise STALL_MS.",+ );++ const withReset = await readsBeforeSettling(Math.floor(withoutReset / 2));+ if (withReset >= withoutReset * 1.25) cleared += 1;+ rounds.push(`${withReset} with a reset against ${withoutReset} without`);+ }- const withReset = await readsBeforeSettling(Math.floor(withoutReset / 2));assert.ok(- withReset >= withoutReset * 1.25,- "a healthy read did not restart the stall window: " +- `${withReset} reads with one, ${withoutReset} without. A run of ` ++ cleared >= 2,+ "a healthy read did not restart the stall window, in " ++ `${cleared} of ${rounds.length} rounds: ${rounds.join("; ")}. A run of ` +"unreadable polls is inheriting the elapsed time of the run before it, so " +"a slow download that keeps reporting progress can still be abandoned.",);