cozyclay 2.1.0 → 3.0.0

package/tools/ardy/run-sequence-on-box.sh

@@ -1,177 +0,0 @@
-#!/usr/bin/env bash
-# Run a multi-prompt ARDY rollout in one remote process. The remote
-# cclay_sequence_generate.py keeps the model loaded and feeds each segment
-# the previous segment's motion tail through init_history_sequence.
-set -euo pipefail
-
-HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-HOST="${CCLAY_ARDY_HOST:-}"
-REMOTE="${CCLAY_ARDY_REPO:-\$HOME/ardy}"
-VENV_PY="${CCLAY_ARDY_VENV:-~/ardy/.venv-cuda/bin/python}"
-ENCODER_URL="${CCLAY_ARDY_ENCODER_URL:-http://127.0.0.1:9550/}"
-FORCE_CPU=0
-
-SSH_OPTS=(-o BatchMode=yes -o ConnectTimeout=10 -o ServerAliveInterval=30 -o ServerAliveCountMax=240)
-SCP_OPTS=(-o BatchMode=yes -o ConnectTimeout=10)
-
-usage() {
-  cat >&2 <<'EOF'
-usage: run-sequence-on-box.sh \
-       --segment "<prompt>" <seconds> [--segment "<prompt>" <seconds> ...] \
-       [--root-2d <frame> <x> <z> <heading|none> ...] \
-       [--seed <N>] [--cpu] [--output <local.npz>] [--dry-run]
-EOF
-  exit 2
-}
-
-SEGMENTS=()
-ROOT2D=()
-SEED=""
-OUTPUT=""
-DRY_RUN=0
-ROOT_MARGIN=""
-CONTACT_THRESHOLD=""
-HISTORY_FRAMES=""
-
-while [[ $# -gt 0 ]]; do
-  case "$1" in
-    --root-2d)
-      [[ $# -ge 5 ]] || { echo "run-sequence-on-box: --root-2d needs FRAME X Z HEADING" >&2; usage; }
-      [[ "$2" =~ ^[0-9]+$ ]] || { echo "run-sequence-on-box: --root-2d frame must be a non-negative integer, got '$2'" >&2; exit 1; }
-      [[ "$3" =~ ^-?[0-9]+([.][0-9]+)?$ && "$4" =~ ^-?[0-9]+([.][0-9]+)?$ ]] || {
-        echo "run-sequence-on-box: --root-2d X Z must be numbers, got '$3' '$4'" >&2; exit 1;
-      }
-      [[ "$5" == "none" || "$5" =~ ^-?[0-9]+([.][0-9]+)?$ ]] || {
-        echo "run-sequence-on-box: --root-2d heading must be a number or 'none', got '$5'" >&2; exit 1;
-      }
-      ROOT2D+=("$2" "$3" "$4" "$5")
-      shift 5 ;;
-    --segment)
-      [[ $# -ge 3 ]] || { echo "run-sequence-on-box: --segment needs PROMPT SECONDS" >&2; usage; }
-      [[ -n "${2//[[:space:]]/}" ]] || { echo "run-sequence-on-box: segment prompt must not be empty" >&2; exit 1; }
-      [[ "$3" =~ ^[0-9]+([.][0-9]+)?$ ]] || { echo "run-sequence-on-box: segment duration must be a number, got '$3'" >&2; exit 1; }
-      SEGMENTS+=("$2" "$3")
-      shift 3 ;;
-    --seed)
-      [[ $# -ge 2 && "$2" =~ ^[0-9]+$ ]] || { echo "run-sequence-on-box: --seed needs a non-negative integer" >&2; usage; }
-      SEED="$2"; shift 2 ;;
-    --cpu)
-      FORCE_CPU=1; shift ;;
-    --root-margin)
-      [[ $# -ge 2 && "$2" =~ ^[0-9]+([.][0-9]+)?$ ]] || { echo "run-sequence-on-box: --root-margin needs a non-negative number" >&2; usage; }
-      ROOT_MARGIN="$2"; shift 2 ;;
-    --history-frames)
-      [[ $# -ge 2 && "$2" =~ ^[0-9]+$ ]] || { echo "run-sequence-on-box: --history-frames needs a non-negative integer" >&2; usage; }
-      HISTORY_FRAMES="$2"; shift 2 ;;
-    --contact-threshold)
-      [[ $# -ge 2 && "$2" =~ ^[0-9]+([.][0-9]+)?$ ]] || { echo "run-sequence-on-box: --contact-threshold needs a non-negative number" >&2; usage; }
-      CONTACT_THRESHOLD="$2"; shift 2 ;;
-    --output)
-      [[ $# -ge 2 ]] || { echo "run-sequence-on-box: --output needs a path" >&2; usage; }
-      OUTPUT="$2"; shift 2 ;;
-    --dry-run)
-      DRY_RUN=1; shift ;;
-    -h|--help)
-      usage ;;
-    *)
-      echo "run-sequence-on-box: unknown option '$1'" >&2
-      usage ;;
-  esac
-done
-
-[[ -n "$HOST" ]] || {
-  echo "run-sequence-on-box: CCLAY_ARDY_HOST is required (for example: user@ardy-host)" >&2
-  exit 2
-}
-
-[[ ${#SEGMENTS[@]} -ge 4 ]] || {
-  echo "run-sequence-on-box: at least two --segment entries are required" >&2
-  exit 1
-}
-
-for ((i = 1; i < ${#SEGMENTS[@]}; i += 2)); do
-  awk -v d="${SEGMENTS[$i]}" 'BEGIN { exit !(d > 0 && d <= 1200) }' || {
-    echo "run-sequence-on-box: segment duration must be > 0 and <= 1200 seconds, got '${SEGMENTS[$i]}'" >&2
-    exit 1
-  }
-done
-
-OUTPUT="${OUTPUT:-$HERE/out/$(date +%s)-sequence.npz}"
-
-if ! ssh "${SSH_OPTS[@]}" "$HOST" ":" </dev/null; then
-  echo "run-sequence-on-box: cannot ssh to ${HOST}" >&2
-  exit 1
-fi
-echo "run-sequence-on-box: ssh to ${HOST} ok"
-
-if ! ssh "${SSH_OPTS[@]}" "$HOST" "test -x ${VENV_PY}" </dev/null; then
-  echo "run-sequence-on-box: missing ${VENV_PY} on ${HOST}" >&2
-  exit 1
-fi
-
-# The generator is cclay-owned: the source of truth is the repo copy next to
-# this script, synced up on every run so the box can never drift behind it.
-if ! ssh "${SSH_OPTS[@]}" "$HOST" "cat > ${REMOTE}/scripts/cclay_sequence_generate.py" \
-  < "${HERE}/cclay_sequence_generate.py"; then
-  echo "run-sequence-on-box: could not sync cclay_sequence_generate.py to ${HOST}" >&2
-  exit 1
-fi
-echo "run-sequence-on-box: sequence generator synced to ${HOST}"
-
-if ! ENCODER_CODE="$(ssh "${SSH_OPTS[@]}" "$HOST" \
-  "curl -s -o /dev/null -w '%{http_code}' -m 10 ${ENCODER_URL}" </dev/null)"; then
-  echo "run-sequence-on-box: could not probe the text encoder" >&2
-  exit 1
-fi
-[[ "$ENCODER_CODE" == "200" ]] || {
-  echo "run-sequence-on-box: text encoder at ${ENCODER_URL} answered '${ENCODER_CODE}', not 200" >&2
-  exit 1
-}
-echo "run-sequence-on-box: text encoder ${ENCODER_URL} responding"
-
-CUDA_ENV=""
-[[ "$FORCE_CPU" -eq 1 ]] && CUDA_ENV='CUDA_VISIBLE_DEVICES="" '
-
-build_remote_cmd() {
-  local tmp_dir="$1" cmd i
-  cmd="cd ${REMOTE} && ${CUDA_ENV}${VENV_PY} scripts/cclay_sequence_generate.py"
-  for ((i = 0; i < ${#SEGMENTS[@]}; i += 2)); do
-    cmd+=" --segment $(printf '%q' "${SEGMENTS[$i]}") $(printf '%q' "${SEGMENTS[$((i + 1))]}")"
-  done
-  for ((i = 0; i < ${#ROOT2D[@]}; i += 4)); do
-    cmd+=" --root-2d $(printf '%q' "${ROOT2D[$i]}") $(printf '%q' "${ROOT2D[$((i + 1))]}")"
-    cmd+=" $(printf '%q' "${ROOT2D[$((i + 2))]}") $(printf '%q' "${ROOT2D[$((i + 3))]}")"
-  done
-  cmd+=" --output $(printf '%q' "${tmp_dir}/out")"
-  [[ -z "$SEED" ]] || cmd+=" --seed $(printf '%q' "$SEED")"
-  [[ -z "$ROOT_MARGIN" ]] || cmd+=" --root-margin $(printf '%q' "$ROOT_MARGIN")"
-  [[ -z "$CONTACT_THRESHOLD" ]] || cmd+=" --contact-threshold $(printf '%q' "$CONTACT_THRESHOLD")"
-  [[ -z "$HISTORY_FRAMES" ]] || cmd+=" --history_frames $(printf '%q' "$HISTORY_FRAMES")"
-  printf '%s' "$cmd"
-}
-
-if [[ "$DRY_RUN" -eq 1 ]]; then
-  echo "run-sequence-on-box: DRY RUN - $(( ${#SEGMENTS[@]} / 2 )) segments in one ARDY process"
-  echo "run-sequence-on-box: would run ssh ${HOST} \"$(build_remote_cmd "<mktemp-dir-on-box>")\""
-  echo "run-sequence-on-box: would pull <mktemp-dir-on-box>/out.npz -> ${OUTPUT}"
-  exit 0
-fi
-
-mkdir -p "$(dirname "$OUTPUT")"
-REMOTE_TMP="$(ssh "${SSH_OPTS[@]}" "$HOST" "mktemp -d" </dev/null)"
-[[ "$REMOTE_TMP" == /* ]] || {
-  echo "run-sequence-on-box: remote mktemp returned unsafe path '${REMOTE_TMP}'" >&2
-  exit 1
-}
-trap 'ssh "${SSH_OPTS[@]}" "$HOST" "rm -rf ${REMOTE_TMP}" </dev/null >/dev/null 2>&1 || true' EXIT
-
-REMOTE_CMD="$(build_remote_cmd "$REMOTE_TMP")"
-echo "run-sequence-on-box: generating $(( ${#SEGMENTS[@]} / 2 )) prompt blocks in one ARDY process ..."
-ssh "${SSH_OPTS[@]}" "$HOST" "$REMOTE_CMD" </dev/null
-
-scp -q "${SCP_OPTS[@]}" "$HOST:${REMOTE_TMP}/out.npz" "$OUTPUT"
-[[ -s "$OUTPUT" ]] || {
-  echo "run-sequence-on-box: pulled ${OUTPUT} is empty" >&2
-  exit 1
-}
-echo "run-sequence-on-box: done - ${OUTPUT} ($(wc -c < "$OUTPUT" | tr -d ' ') bytes)"

package/tools/ardy/runners/gvhmr-worker.mjs

@@ -1,206 +0,0 @@
-/** Serial, bridge-owned SSH worker. Cached weights are on CPU between jobs.
- * Disconnect/timeout kills the worker; the next request starts a clean one.
- */
-import { spawn } from "node:child_process";
-import { createHash, randomUUID } from "node:crypto";
-import { readFileSync } from "node:fs";
-import { fileURLToPath } from "node:url";
-import { killGroup, streamLines, track } from "./proc.mjs";
-
-export class PersistentGvhmrWorker {
-	constructor({ prepare = async () => {}, start, idleMs = 10 * 60_000 }) {
-		this.prepare = prepare;
-		this.start = start;
-		this.idleMs = idleMs;
-		this.child = null;
-		this.ready = null;
-		this.active = null;
-		this.idle = null;
-	}
-
-	run(request, { signal, onLine, timeoutMs = 30 * 60_000 } = {}) {
-		if (signal?.aborted) return Promise.reject(new Error("extract-cancelled"));
-		if (this.active) return Promise.reject(new Error("extract-worker-busy"));
-		clearTimeout(this.idle);
-		return new Promise((resolve, reject) => {
-			const ctx = { request: { ...request, id: randomUUID() }, resolve, reject, signal, onLine,
-				started: performance.now(), controller: new AbortController(), startupSeconds: 0 };
-			this.active = ctx;
-			ctx.abort = () => this.stop("extract-cancelled");
-			signal?.addEventListener("abort", ctx.abort, { once: true });
-			ctx.timer = setTimeout(() => this.stop("extract-timeout"), timeoutMs);
-			this.begin(ctx).catch((err) => {
-				if (this.active === ctx) this.stop(`extract-worker-start: ${err.message}`);
-			});
-		});
-	}
-
-	async begin(ctx) {
-		if (this.child && this.ready) return this.dispatch(ctx);
-		await this.prepare(ctx.controller.signal);
-		if (this.active !== ctx) return; // cancellation while deploying must not spawn a late worker
-		const child = this.start();
-		this.child = child;
-		track(child);
-		child.stdin.on("error", (err) => {
-			if (this.child === child) this.stop(`extract-worker-stdin: ${err.message}`);
-		});
-		streamLines(child.stdout, (line) => {
-			if (this.child !== child) return;
-			let msg;
-			try { msg = JSON.parse(line); } catch { return this.stop("extract-worker-protocol"); }
-			if (msg.event === "ready" && msg.protocol === 1 && !this.ready) {
-				this.ready = msg;
-				if (this.active) {
-					this.active.startupSeconds = msg.startupSeconds;
-					this.dispatch(this.active);
-				}
-				return;
-			}
-			const active = this.active;
-			if (!active || !active.sent || msg.id !== active.request.id) return this.stop("extract-worker-protocol");
-			if (msg.event === "error") return this.stop(`extract-worker-failed: ${msg.message}`);
-			if (msg.event !== "done") return this.stop("extract-worker-protocol");
-			this.finish(null, { ...msg.performance, workerStartupSeconds: active.startupSeconds,
-				workerWallSeconds: (performance.now() - active.started) / 1000, runnerSha256: this.ready.runnerSha256 });
-		});
-		streamLines(child.stderr, (line) => {
-			if (this.child === child) this.active?.onLine?.(line);
-		});
-		child.once("error", (err) => {
-			if (this.child === child) this.stop(`extract-worker-spawn: ${err.message}`);
-		});
-		child.once("close", (code) => {
-			if (this.child !== child) return;
-			this.child = this.ready = null;
-			this.finish(new Error(`extract-worker-exited: ${code}`));
-		});
-	}
-
-	dispatch(ctx) {
-		if (ctx.sent || this.active !== ctx) return;
-		ctx.sent = true;
-		this.child.stdin.write(`${JSON.stringify(ctx.request)}\n`);
-	}
-
-	finish(error, result) {
-		const ctx = this.active;
-		this.active = null;
-		if (ctx) {
-			clearTimeout(ctx.timer);
-			ctx.signal?.removeEventListener("abort", ctx.abort);
-			ctx.controller.abort();
-			if (error) ctx.reject(error);
-			else ctx.resolve(result);
-		}
-		if (this.child) {
-			clearTimeout(this.idle);
-			this.idle = setTimeout(() => this.stop(), this.idleMs);
-			this.idle.unref();
-		}
-	}
-
-	stop(reason = "extract-worker-stopped") {
-		clearTimeout(this.idle);
-		const child = this.child;
-		this.child = this.ready = null;
-		if (child) {
-			child.stdin.end(); // remote EOF terminates active GPU work as well
-			killGroup(child);
-		}
-		this.finish(new Error(reason));
-	}
-}
-
-function setup(command, args, signal) {
-	return new Promise((resolve, reject) => {
-		if (signal.aborted) return reject(new Error("extract-cancelled"));
-		const child = spawn(command, args, { detached: true, stdio: ["ignore", "ignore", "pipe"] });
-		track(child);
-		let error = "";
-		const abort = () => killGroup(child);
-		signal.addEventListener("abort", abort, { once: true });
-		child.stderr.on("data", (chunk) => { error = (error + chunk).slice(-4096); });
-		child.once("error", reject);
-		child.once("close", (code) => {
-			signal.removeEventListener("abort", abort);
-			if (code === 0 && !signal.aborted) resolve();
-			else reject(new Error(error.trim() || `${command} exited ${code}`));
-		});
-	});
-}
-
-// CozyClay's mocap input is the coloured mannequin.  Palette segmentation is
-// therefore a contract, rather than a selectable heuristic: keeping this
-// list to one value prevents an environment override from silently routing a
-// take through YOLO or the auto detector.
-const DETECTORS = ["palette"];
-// hybrid (#380): ViTPose where the render kept the limb part, palette-derived
-// where it lost one (a prop repainting the hand). The runner's `auto` on a
-// palette clip is hybrid.
-const KEYPOINTS = ["vitpose", "palette", "hybrid", "auto"];
-// The runner's temporal smoother (quaternion gaussian over the SMPL params),
-// in frames. GVHMR has no output filter; its runner default of 1.2 measured
-// 4.61 mm/f² jitter on a 24 fps walk against 3.39 at 3.0 with the stride
-// intact (#380). Sent on both extraction paths so they extract the same take.
-export const GVHMR_SMOOTH_SIGMA = 3;
-
-/** Runner CLI flags shared by the one-shot ssh command and the worker.
- *  Pure so the flag wiring is testable without a box: the part-coloured
- *  mannequin (#137) is invisible to YOLO's person class (28/124 frames at
- *  conf 0.5 against 481/481 for a real person), so the runner always uses its
- *  palette detector.
- *
- *  `keypoints` picks what fills GVHMR's kp2d observation (#180). ViTPose has
- *  to infer joints a flat-coloured render gives it no cue for — measured on
- *  h3_warm_s7 its median body joint sits 10.4 % of bbox height off the
- *  palette joints, the shoulders 30-65 % — while the mannequin's per-segment
- *  hues put every joint exactly where two segment masks meet. `auto` follows
- *  the detector: palette keypoints for a palette-detected clip, ViTPose for
- *  real footage, which is the only place ViTPose is the better estimate.
- */
-export function gvhmrRunnerArgs({ staticCam = true, fMm = null, detector = "palette", keypoints = "auto", smoothSigma = GVHMR_SMOOTH_SIGMA } = {}) {
-	const args = [];
-	if (staticCam) args.push("--static-cam");
-	if (fMm != null) args.push("--f-mm", String(Math.trunc(fMm)));
-	// Always send the flag.  Omitting it would let the remote runner's own
-	// default (`auto`) bypass the palette contract.
-	args.push("--detector", "palette");
-	args.push("--keypoints", KEYPOINTS.includes(keypoints) ? keypoints : "auto");
-	args.push("--smooth-sigma", String(Number.isFinite(smoothSigma) && smoothSigma > 0 ? smoothSigma : GVHMR_SMOOTH_SIGMA));
-	return args;
-}
-
-/** CCLAY_EXTRACT_DETECTOR is retained for deployment compatibility, but cannot
- *  change the detector. Every value resolves to the palette contract. */
-export function gvhmrDetectorFromEnv(env = process.env) {
-	return "palette";
-}
-
-/** CCLAY_EXTRACT_KEYPOINTS as the runner understands it, degrading unknown
- *  values to `auto` for compatibility with the existing keypoint A/B switch. */
-export function gvhmrKeypointsFromEnv(env = process.env) {
-	const value = env.CCLAY_EXTRACT_KEYPOINTS?.trim().toLowerCase() || "auto";
-	return KEYPOINTS.includes(value) ? value : "auto";
-}
-
-const clients = new Map();
-export function gvhmrWorker({ host, sshOptions, scpOptions }) {
-	const key = JSON.stringify([host, sshOptions, scpOptions]);
-	if (clients.has(key)) return clients.get(key);
-	const files = ["cclay_gvhmr_worker.py", "gvhmr_fastpath.py", "gvhmr_trajectory.py"].map((name) => fileURLToPath(new URL(`../${name}`, import.meta.url)));
-	const hash = createHash("sha256");
-	for (const file of files) hash.update(readFileSync(file));
-	const directory = `/tmp/cozyclay-gvhmr-worker-${hash.digest("hex").slice(0, 24)}`;
-	const client = new PersistentGvhmrWorker({
-		prepare: async (signal) => {
-			await setup("ssh", [...sshOptions, host, `umask 077 && mkdir -p '${directory}'`], signal);
-			await setup("scp", [...scpOptions, ...files, `${host}:${directory}/`], signal);
-		},
-		start: () => spawn("ssh", [...sshOptions, "-o", "ServerAliveInterval=15", "-o", "ServerAliveCountMax=2", host,
-			`cd ~/cclay-ingest/GVHMR && exec .venv/bin/python -u '${directory}/cclay_gvhmr_worker.py'`],
-			{ detached: true, stdio: ["pipe", "pipe", "pipe"] }),
-	});
-	clients.set(key, client);
-	return client;
-}

package/tools/ardy/runners/index.mjs

@@ -1,18 +0,0 @@
-/**
- * Runner selection for the motion generation bridge.
- *
- * The local Studio has one supported backend: Kimodo. The former ARDY
- * local/remote selector lived here, which made an unrelated legacy runtime
- * part of every bridge launch. Keeping selection in one place still gives the
- * bridge a small, testable seam while refusing stale ARDY environment values.
- */
-
-import { createKimodoRunner } from "../../kimodo/runner.mjs";
-
-export function createRunner() {
-	const backend = (process.env.CCLAY_MOTION_BACKEND || "").trim().toLowerCase();
-	if (backend && backend !== "kimodo") {
-		throw new Error(`unknown CCLAY_MOTION_BACKEND "${backend}" (expected "kimodo")`);
-	}
-	return createKimodoRunner();
-}