Deploy simready-assets run 20261020-a3cb7f2 (GPU variant)

Dockerfile

@@ -1,4 +1,4 @@
-# deploy: 20261019-9f72902 (simready-assets)
+# deploy: 20261020-a3cb7f2 (simready-assets)
 # GPU variant of the single-container MVP image for Hugging Face Spaces
 # (Docker SDK): everything docker/Dockerfile.hf-space has (OpenSearch + Neo4j
 # + Redis + deepsearch-api + asset-graph-service + nginx under supervisord),
@@ -349,12 +349,12 @@ ENV HF_STORAGE_API_TOKEN= \
     ASSERT_ADMIN_USER=false \
     USDSEARCH_VISION_VALIDATION_ENABLED=false \
     USDSEARCH_LLM_PARSING_ENABLED=false \
-    # Validation limits are per deepsearch-api worker: 16 checks in flight and a 125k-verdict cache in
-    # each of the 4 workers (DEEPSEARCH_API_WORKERS below), 64 and 500k in all. One search's page runs
-    # at most 16 at once anyway (usdsearch's BATCH_CONCURRENCY); with 8, a 50-check page took ~3 s on
-    # the endpoint, with 16 ~1.5 s (2026-10-15 benchmark on the endpoint's own vLLM args).
+    # Validation limits are per deepsearch-api process: 16 checks in flight and an 83k-verdict cache in
+    # each of the 6 processes (DEEPSEARCH_API_PROCESSES below), 96 and ~500k in all (Redis keeps every verdict
+    # anyway). One search's page runs at most 16 at once (usdsearch's BATCH_CONCURRENCY); with 8, a 50-check
+    # page took ~3 s on the endpoint, with 16 ~1.5 s (2026-10-15 benchmark on the endpoint's own vLLM args).
     USDSEARCH_VISION_VALIDATION_MAX_CONCURRENT=16 \
-    USDSEARCH_VISION_VALIDATION_CACHE_SIZE=125000 \
+    USDSEARCH_VISION_VALIDATION_CACHE_SIZE=83333 \
     USDSEARCH_LLM_PARSING_FIELDS_FILEPATH=/app/search_fields.yaml \
     USDSEARCH_LLM_PARSING_PROMPT_FILEPATH=/app/llm_parse_prompt.txt \
     # Business telemetry — same as docker/Dockerfile.hf-space, see its comment.
@@ -384,11 +384,14 @@ ENV HF_STORAGE_API_TOKEN= \
     SIGLIP2_TRITON_SERVER_URL=localhost:8001 \
     TRITON_SERVER_URL=localhost:8001 \
     SIGLIP2_TRITON_SERVER_SSL=false \
-    # deepsearch-api workers — see docker/Dockerfile.hf-space's comment. 4, not
-    # 1, for the production t4-medium (8 vCPU / 30 GB): measured best there
-    # (docs/search-performance.md) — 54 vs 26 req/s at 50 concurrent users;
-    # 10 workers starve OpenSearch of CPU. Override with a Space Variable.
-    DEEPSEARCH_API_WORKERS=4 \
+    # deepsearch-api processes, ONE uvicorn worker each, behind nginx's least_conn
+    # upstream (supervisord-gpu.conf, nginx.conf.template). 6 for a10g-large
+    # (12 vCPU / 46 GB): 6, 8 and 12 carried the same load in the lab (docs/
+    # search-performance.md), 6 costs the least memory. Must equal supervisord-gpu.conf's
+    # numprocs: nginx renders its upstream from this, so never set it as a Space
+    # Variable. DEEPSEARCH_API_WORKERS is no longer read (a leftover Space Variable
+    # of that name does nothing).
+    DEEPSEARCH_API_PROCESSES=6 \
     PORT=7860
 
 ARG RUN_ID

README.md

@@ -5,7 +5,7 @@ colorFrom: green
 colorTo: blue
 sdk: docker
 app_port: 7860
-hardware: t4-medium
+hardware: a10g-large
 ---
 
 # SimReady Central USD Search
@@ -13,7 +13,7 @@ hardware: t4-medium
 SimReady Central USD Search — single-container demo (OpenSearch + Neo4j +
 Redis + deepsearch-api + asset-graph-service, SigLIP2 embeddings via a local
 siglip2-triton process — see docker/Dockerfile.hf-space-gpu). Runs on a GPU
-hardware tier (t4-medium), unlike the free-tier
+hardware tier (a10g-large), unlike the free-tier
 docker/Dockerfile.hf-space variant, which routes SigLIP2 to NVCF instead.
 The index itself is pulled from the artifact S3 bucket at container
 startup, not baked in. Deployed by

docker/hf-space/nginx.conf.template

@@ -1,10 +1,11 @@
 # Friendly external paths remapped to deepsearch-api's/asset-graph-service's
 # actual routes — same mapping as
 # usdsearch/helm/usdsearch/templates/api_gateway_config_map.yaml.
-# deepsearch-api listens on 127.0.0.1:8000, asset-graph-service on
-# 127.0.0.1:8010 (see docker/hf-space/supervisord.conf) — nginx is the sole
-# process bound to the externally proxied port, so it can front every route
-# with these rewrites plus CORS.
+# deepsearch-api runs as single-worker processes on 127.0.0.1:8101, 8102, ...
+# (the `deepsearch` upstream below; how many: DEEPSEARCH_API_PROCESSES, see
+# docker/hf-space/supervisord-gpu.conf), asset-graph-service on
+# 127.0.0.1:8010 — nginx is the sole process bound to the externally proxied
+# port, so it can front every route with these rewrites plus CORS.
 #
 # The Explorer UI is served at `/` (root), not proxied to deepsearch-api at
 # all — served directly from the static bundle baked into this image at
@@ -103,6 +104,14 @@ map $http_accept_encoding $slim_accepts_br {
     "~*(^|,)\s*br\s*(;|,|$)"                               1;
 }
 # <<< simready-central slim explorer (http) <<<
+# deepsearch-api's processes, one uvicorn worker each. least_conn, not `uvicorn --workers N`: with N workers on one
+# socket the kernel hands nearly every connection to 2-3 of them, and the rest of the CPU sits idle (lab load test
+# 2026-10-15, docs/search-performance.md). __DEEPSEARCH_API_SERVERS__ is rendered at container start from
+# DEEPSEARCH_API_PROCESSES (the nginx program in supervisord*.conf).
+upstream deepsearch {
+    least_conn;
+    __DEEPSEARCH_API_SERVERS__
+}
 server {
     listen __PORT__;
 
@@ -123,25 +132,25 @@ server {
     location /search {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/search(.*)$ /v2/deepsearch/search$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
     }
 
     location /search_hybrid {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/search_hybrid(.*)$ /v3/deepsearch/search$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
     }
 
     location /vlm_validate/search_result {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/vlm_validate/search_result(.*)$ /v3/deepsearch/vlm_validate/search_result$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
     }
 
     location /llm_parse/ {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/llm_parse/(.*)$ /v3/deepsearch/llm_parse/$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
     }
 
     # package-download (127.0.0.1:8020, pipeline/package_download/), not deepsearch-api: it answers
@@ -187,7 +196,7 @@ server {
     location /images {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/images(.*)$ /v3/images$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
     }
 
     location /static/ {
@@ -227,8 +236,8 @@ server {
     # below serves the Explorer SPA instead of proxying everything
     # unmatched. /docs and /redoc are prefix matches, not exact: Swagger
     # UI/ReDoc can request sub-resources under their own path.
-    location = /health { proxy_pass http://127.0.0.1:8000; }
-    location = /openapi.json { proxy_pass http://127.0.0.1:8000; }
+    location = /health { proxy_pass http://deepsearch; }
+    location = /openapi.json { proxy_pass http://deepsearch; }
     # /metrics is NOT deepsearch-api's (127.0.0.1:8000/metrics stays private to
     # the container): it is pipeline/cache_metrics.py's page, an allow-list of
     # cache hit/miss counters (supervisord program cache-metrics). Not up yet:
@@ -241,8 +250,8 @@ server {
         default_type application/json;
         return 503 '{"detail":"cache metrics are not running"}';
     }
-    location /docs { proxy_pass http://127.0.0.1:8000; }
-    location /redoc { proxy_pass http://127.0.0.1:8000; }
+    location /docs { proxy_pass http://deepsearch; }
+    location /redoc { proxy_pass http://deepsearch; }
 
     # The Explorer SPA (see header comment) at root: try_files serves any of
     # its own literal files (index.html, ds.png, favicon.ico, manifest.json,
@@ -389,7 +398,7 @@ server {
     location = /search/stats/usd_properties {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/search(.*)$ /v2/deepsearch/search$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
         proxy_ignore_headers Cache-Control Expires;
         proxy_cache slim_stats_cache;
         proxy_cache_key "$request_uri";
@@ -415,7 +424,7 @@ server {
     location = /images {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/images(.*)$ /v3/images$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
         proxy_cache slim_images_cache;
         proxy_cache_key "$request_uri";
         proxy_ignore_headers Cache-Control Expires;
@@ -446,7 +455,7 @@ server {
     location = /llm_parse/query {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^/llm_parse/(.*)$ /v3/deepsearch/llm_parse/$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
         proxy_cache slim_parse_cache;
         proxy_cache_key "$slim_today|$request_uri|$request_body";
         proxy_cache_bypass $slim_not_post $request_body_file;
@@ -473,7 +482,7 @@ server {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         if ($slim_count_probe) { rewrite ^ /_slim_count_probe last; }
         rewrite ^/search_hybrid(.*)$ /v3/deepsearch/search$1 break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
     }
     # The Explorer's background count probes (limit-1 and file-name searches for its totals; SimReady Central's
     # gateway caches them the same way): their own zone, 30 min, stale while one request refreshes an entry. Never a
@@ -481,7 +490,7 @@ server {
     location = /_slim_count_probe {
         internal;
         rewrite ^ /v3/deepsearch/search break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
         proxy_cache slim_count_cache;
         proxy_cache_key "$request_uri|$request_body";
         proxy_cache_bypass $slim_not_post $request_body_file $slim_probe_skip;
@@ -533,7 +542,7 @@ server {
     location = /vlm_validate/batch_stream {
         if ($request_method = OPTIONS) { add_header Content-Length 0 always; add_header Content-Type "text/plain" always; return 204; }
         rewrite ^ /v3/deepsearch/vlm_validate/search_result_batch break;
-        proxy_pass http://127.0.0.1:8000;
+        proxy_pass http://deepsearch;
         proxy_buffering off;
         proxy_read_timeout 120s;
         gzip off;
@@ -616,3 +625,21 @@ server {
         try_files $uri $uri/ /index.html;
     }
 }
+# The API's old address, for the in-container callers that use it directly (package-download's fallback to the
+# API's own /v3/download, its size and spell-vocab passes): spread over the same processes. Loopback only.
+server {
+    listen 127.0.0.1:8000;
+    access_log off;
+    client_max_body_size 0;
+    proxy_http_version 1.1;
+    proxy_set_header Connection "";
+    proxy_set_header Host $host;
+    proxy_buffering off;
+    proxy_request_buffering off;
+    proxy_connect_timeout 5s;
+    proxy_send_timeout 600s;
+    proxy_read_timeout 600s;
+    location / {
+        proxy_pass http://deepsearch;
+    }
+}

docker/hf-space/supervisord-gpu.conf

@@ -159,19 +159,45 @@ stdout_logfile_maxbytes=0
 stderr_logfile=/dev/stderr
 stderr_logfile_maxbytes=0
 
+[program:deepsearch-seed]
+; Runs once per container start, before any deepsearch-api process: waits on
+; siglip2-triton's own readiness (via its remapped HTTP port, 8003 — see
+; [program:siglip2-triton] above), runs pipeline/hf_space_seed.py, then
+; empties PROMETHEUS_MULTIPROC_DIR, where the processes share their metrics
+; (a counter is the sum over all of them; [program:cache-metrics] reads its
+; cache counters there), and leaves /tmp/deepsearch-seed.done for them. Not
+; in deepsearch-api's own command: each of its processes would wipe the
+; directory the others write to.
+command=/usr/bin/env -u PYTHONPATH /bin/sh -c "rm -f /tmp/deepsearch-seed.done && until curl -sf http://localhost:8003/v2/health/ready >/dev/null 2>&1; do sleep 2; done; cd /app/hf-space-pipeline && /app/.venv/bin/python -m pipeline.hf_space_seed && rm -rf /tmp/deepsearch-api-prometheus && mkdir -p /tmp/deepsearch-api-prometheus && touch /tmp/deepsearch-seed.done"
+directory=/app
+user=1000
+autorestart=unexpected
+exitcodes=0
+startsecs=0
+; Started before the API processes (lower priority), which also wait 3 s before looking for
+; the marker: a /tmp kept across a container restart can't let them skip this run's seed.
+priority=10
+stdout_logfile=/dev/stdout
+stdout_logfile_maxbytes=0
+stderr_logfile=/dev/stderr
+stderr_logfile_maxbytes=0
+
 [program:deepsearch-api]
-; Waits on siglip2-triton's own readiness (via its remapped HTTP port,
-; 8003 — see [program:siglip2-triton] above) in addition to running
-; pipeline/hf_space_seed.py, same "don't exec uvicorn until every backend
-; this process depends on is actually ready" rule the cpu-basic/NVCF
-; variant already follows for OpenSearch.
-; PROMETHEUS_MULTIPROC_DIR (emptied at every start): the workers share their
-; metrics through files there, so a counter is the sum over all workers;
-; [program:cache-metrics] reads its cache counters from them.
+; 6 processes with ONE uvicorn worker each, on 127.0.0.1:8101-8106, behind
+; nginx's least_conn upstream (nginx.conf.template). Not `uvicorn --workers 6`:
+; there the kernel hands nearly every connection to 2-3 workers and most of
+; the CPU idles (docs/search-performance.md). numprocs is read from
+; DEEPSEARCH_API_PROCESSES (Dockerfile.hf-space-gpu), the same value nginx renders its
+; upstream from, so the two can't differ (a Space Variable changes both). Supervisord
+; restarts one process that dies without touching the others. Each waits for
+; [program:deepsearch-seed].
 ; USDSEARCH_LLM_PARSING_CACHE_SIZE 0 (a Space variable overrides it): no
-; per-worker parse LRU. Its key has no date, so after midnight it answered
+; per-process parse LRU. Its key has no date, so after midnight it answered
 ; "made last week" with yesterday's parse; nginx caches parses per day instead.
-command=/usr/bin/env -u PYTHONPATH /bin/sh -c "until curl -sf http://localhost:8003/v2/health/ready >/dev/null 2>&1; do sleep 2; done; cd /app/hf-space-pipeline && /app/.venv/bin/python -m pipeline.hf_space_seed && cd /app && rm -rf /tmp/deepsearch-api-prometheus && mkdir -p /tmp/deepsearch-api-prometheus && exec env PROMETHEUS_MULTIPROC_DIR=/tmp/deepsearch-api-prometheus USDSEARCH_LLM_PARSING_CACHE_SIZE=\"${USDSEARCH_LLM_PARSING_CACHE_SIZE:-0}\" /app/.venv/bin/python -m uvicorn deepsearch_api.main:app --host 127.0.0.1 --port 8000 --workers \"${DEEPSEARCH_API_WORKERS:-1}\""
+process_name=%(program_name)s_%(process_num)02d
+numprocs=%(ENV_DEEPSEARCH_API_PROCESSES)s
+numprocs_start=1
+command=/usr/bin/env -u PYTHONPATH /bin/sh -c "sleep 3; until [ -f /tmp/deepsearch-seed.done ]; do sleep 2; done; cd /app && exec env PROMETHEUS_MULTIPROC_DIR=/tmp/deepsearch-api-prometheus USDSEARCH_LLM_PARSING_CACHE_SIZE=\"${USDSEARCH_LLM_PARSING_CACHE_SIZE:-0}\" /app/.venv/bin/python -m uvicorn deepsearch_api.main:app --host 127.0.0.1 --port 81%(process_num)02d --workers 1"
 directory=/app
 user=1000
 autorestart=true
@@ -200,10 +226,12 @@ stderr_logfile=/dev/stderr
 stderr_logfile_maxbytes=0
 
 [program:nginx]
-; See docker/hf-space/supervisord.conf's own copy (identical command). Never put
+; See docker/hf-space/supervisord.conf's own copy (identical command): it waits
+; for all DEEPSEARCH_API_PROCESSES API processes (after ~90 s, for at least one: nginx's
+; upstream skips a process that doesn't answer) and renders the upstream from the same list. Never put
 ; a space before a ";" or "#" in a command: supervisord reads it as an inline
 ; comment and cuts the command there (2026-10-15: nginx FATAL on prod).
-command=/bin/sh -c "until curl -sf http://localhost:8000/health >/dev/null 2>&1 && curl -sf http://localhost:8010/docs >/dev/null 2>&1; do sleep 2; done; dl=false; case '%(ENV_DOWNLOAD_ENABLED)s' in [Tt]rue|1) if [ -f /app/hf-space-pipeline/.build/restore/hf-storage-api-catalog.yaml ]; then dl=true; else echo 'nginx: no hf-storage-api catalog, rendering download:false' >&2; fi;; esac; vf=glb; case $VIEWER_FORMAT in glb|usd|auto) vf=$VIEWER_FORMAT;; *) [ -z \"$VIEWER_FORMAT\" ] || echo 'nginx: VIEWER_FORMAT is not glb, usd or auto: rendering viewerFormat glb' >&2;; esac; sed -e 's/__PORT__/%(ENV_PORT)s/' -e 's|__CORPUS_NAME__|%(ENV_CORPUS_NAME)s|g' -e 's/__DOWNLOAD_ENABLED__/'$dl'/' -e 's|__SUBMIT_CONTENT_URL__|%(ENV_SUBMIT_CONTENT_URL)s|g' -e 's/__VIEWER_FORMAT__/'$vf'/' /etc/nginx/hf-space.conf.template > /etc/nginx/conf.d/hf-space.conf && exec nginx -g 'daemon off;'"
+command=/bin/sh -c "n=${DEEPSEARCH_API_PROCESSES:-1}; ports=; ups=; i=1; while [ $i -le $n ]; do ports=\"$ports $((8100+i))\"; ups=\"$ups server 127.0.0.1:$((8100+i));\"; i=$((i+1)); done; t=0; until curl -sf http://localhost:8010/docs >/dev/null 2>&1 && { (for p in $ports; do curl -sf http://localhost:$p/health >/dev/null 2>&1 || exit 1; done) || { [ $t -ge 90 ] && (for p in $ports; do curl -sf http://localhost:$p/health >/dev/null 2>&1 && exit 0; done; exit 1); }; }; do sleep 2; t=$((t+2)); done; dl=false; case '%(ENV_DOWNLOAD_ENABLED)s' in [Tt]rue|1) if [ -f /app/hf-space-pipeline/.build/restore/hf-storage-api-catalog.yaml ]; then dl=true; else echo 'nginx: no hf-storage-api catalog, rendering download:false' >&2; fi;; esac; vf=glb; case $VIEWER_FORMAT in glb|usd|auto) vf=$VIEWER_FORMAT;; *) [ -z \"$VIEWER_FORMAT\" ] || echo 'nginx: VIEWER_FORMAT is not glb, usd or auto: rendering viewerFormat glb' >&2;; esac; sed -e 's/__PORT__/%(ENV_PORT)s/' -e 's|__CORPUS_NAME__|%(ENV_CORPUS_NAME)s|g' -e 's/__DOWNLOAD_ENABLED__/'$dl'/' -e 's|__SUBMIT_CONTENT_URL__|%(ENV_SUBMIT_CONTENT_URL)s|g' -e 's/__VIEWER_FORMAT__/'$vf'/' -e \"s/__DEEPSEARCH_API_SERVERS__/$ups/\" /etc/nginx/hf-space.conf.template > /etc/nginx/conf.d/hf-space.conf && exec nginx -g 'daemon off;'"
 user=root
 autorestart=true
 startsecs=3

docker/hf-space/supervisord.conf

@@ -171,9 +171,11 @@ stderr_logfile=/dev/stderr
 stderr_logfile_maxbytes=0
 
 [program:deepsearch-api]
-; Bound to localhost only — nginx (below) is the sole process bound to the
-; externally proxied port, so it can front every route with the friendly-URL
-; rewrites and CORS headers in nginx.conf.template.
+; Bound to localhost only, on 127.0.0.1:8101: the one server of nginx's
+; `deepsearch` upstream in this image (DEEPSEARCH_API_PROCESSES=1; the GPU
+; image runs several, see supervisord-gpu.conf). nginx (below) is the sole
+; process bound to the externally proxied port, so it can front every route
+; with the friendly-URL rewrites and CORS headers in nginx.conf.template.
 ;
 ; Runs pipeline/hf_space_seed.py first and only execs uvicorn once it
 ; succeeds: that script waits for opensearch (started in parallel by
@@ -190,7 +192,7 @@ stderr_logfile_maxbytes=0
 ; USDSEARCH_LLM_PARSING_CACHE_SIZE 0 (a Space variable overrides it): no
 ; per-worker parse LRU. Its key has no date, so after midnight it answered
 ; "made last week" with yesterday's parse; nginx caches parses per day instead.
-command=/bin/sh -c "cd /app/hf-space-pipeline && /app/.venv/bin/python -m pipeline.hf_space_seed && cd /app && rm -rf /tmp/deepsearch-api-prometheus && mkdir -p /tmp/deepsearch-api-prometheus && exec env PROMETHEUS_MULTIPROC_DIR=/tmp/deepsearch-api-prometheus USDSEARCH_LLM_PARSING_CACHE_SIZE=\"${USDSEARCH_LLM_PARSING_CACHE_SIZE:-0}\" /app/.venv/bin/python -m uvicorn deepsearch_api.main:app --host 127.0.0.1 --port 8000 --workers \"${DEEPSEARCH_API_WORKERS:-1}\""
+command=/bin/sh -c "cd /app/hf-space-pipeline && /app/.venv/bin/python -m pipeline.hf_space_seed && cd /app && rm -rf /tmp/deepsearch-api-prometheus && mkdir -p /tmp/deepsearch-api-prometheus && exec env PROMETHEUS_MULTIPROC_DIR=/tmp/deepsearch-api-prometheus USDSEARCH_LLM_PARSING_CACHE_SIZE=\"${USDSEARCH_LLM_PARSING_CACHE_SIZE:-0}\" /app/.venv/bin/python -m uvicorn deepsearch_api.main:app --host 127.0.0.1 --port 8101 --workers 1"
 directory=/app
 user=1000
 autorestart=true
@@ -240,11 +242,15 @@ stderr_logfile_maxbytes=0
 ; %(ENV_PORT)s is substituted by supervisord itself, so no extra templating
 ; tool (envsubst etc.) is needed in the image.
 ;
+; Waits for asset-graph-service and every deepsearch-api process
+; (127.0.0.1:8101 .. 8100+DEEPSEARCH_API_PROCESSES), and renders the
+; template's `deepsearch` upstream from that same list.
+;
 ; Never put a space before a ";" or "#" in a command: supervisord's parser
 ; reads " ;" and " #" as the start of an inline comment and cuts the command
 ; there (2026-10-15: a " ;;" case arm left nginx FATAL, "No closing
 ; quotation"). tests/test_hf_space_runtime.py parses with supervisor itself.
-command=/bin/sh -c "until curl -sf http://localhost:8000/health >/dev/null 2>&1 && curl -sf http://localhost:8010/docs >/dev/null 2>&1; do sleep 2; done; dl=false; case '%(ENV_DOWNLOAD_ENABLED)s' in [Tt]rue|1) if [ -f /app/hf-space-pipeline/.build/restore/hf-storage-api-catalog.yaml ]; then dl=true; else echo 'nginx: no hf-storage-api catalog, rendering download:false' >&2; fi;; esac; vf=glb; case $VIEWER_FORMAT in glb|usd|auto) vf=$VIEWER_FORMAT;; *) [ -z \"$VIEWER_FORMAT\" ] || echo 'nginx: VIEWER_FORMAT is not glb, usd or auto: rendering viewerFormat glb' >&2;; esac; sed -e 's/__PORT__/%(ENV_PORT)s/' -e 's|__CORPUS_NAME__|%(ENV_CORPUS_NAME)s|g' -e 's/__DOWNLOAD_ENABLED__/'$dl'/' -e 's|__SUBMIT_CONTENT_URL__|%(ENV_SUBMIT_CONTENT_URL)s|g' -e 's/__VIEWER_FORMAT__/'$vf'/' /etc/nginx/hf-space.conf.template > /etc/nginx/conf.d/hf-space.conf && exec nginx -g 'daemon off;'"
+command=/bin/sh -c "n=${DEEPSEARCH_API_PROCESSES:-1}; ports=; ups=; i=1; while [ $i -le $n ]; do ports=\"$ports $((8100+i))\"; ups=\"$ups server 127.0.0.1:$((8100+i));\"; i=$((i+1)); done; t=0; until curl -sf http://localhost:8010/docs >/dev/null 2>&1 && { (for p in $ports; do curl -sf http://localhost:$p/health >/dev/null 2>&1 || exit 1; done) || { [ $t -ge 90 ] && (for p in $ports; do curl -sf http://localhost:$p/health >/dev/null 2>&1 && exit 0; done; exit 1); }; }; do sleep 2; t=$((t+2)); done; dl=false; case '%(ENV_DOWNLOAD_ENABLED)s' in [Tt]rue|1) if [ -f /app/hf-space-pipeline/.build/restore/hf-storage-api-catalog.yaml ]; then dl=true; else echo 'nginx: no hf-storage-api catalog, rendering download:false' >&2; fi;; esac; vf=glb; case $VIEWER_FORMAT in glb|usd|auto) vf=$VIEWER_FORMAT;; *) [ -z \"$VIEWER_FORMAT\" ] || echo 'nginx: VIEWER_FORMAT is not glb, usd or auto: rendering viewerFormat glb' >&2;; esac; sed -e 's/__PORT__/%(ENV_PORT)s/' -e 's|__CORPUS_NAME__|%(ENV_CORPUS_NAME)s|g' -e 's/__DOWNLOAD_ENABLED__/'$dl'/' -e 's|__SUBMIT_CONTENT_URL__|%(ENV_SUBMIT_CONTENT_URL)s|g' -e 's/__VIEWER_FORMAT__/'$vf'/' -e \"s/__DEEPSEARCH_API_SERVERS__/$ups/\" /etc/nginx/hf-space.conf.template > /etc/nginx/conf.d/hf-space.conf && exec nginx -g 'daemon off;'"
 user=root
 autorestart=true
 startsecs=3

services/explorer/dist/static/js/main.01872f5b.js.LICENSE.txt

@@ -1,39 +0,0 @@
-/**
- * @license React
- * react-dom.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */
-
-/**
- * @license React
- * react-jsx-runtime.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */
-
-/**
- * @license React
- * react.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */
-
-/**
- * @license React
- * scheduler.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */

services/explorer/dist/static/js/main.0ed32d3d.js.LICENSE.txt

@@ -1,39 +0,0 @@
-/**
- * @license React
- * react-dom.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */
-
-/**
- * @license React
- * react-jsx-runtime.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */
-
-/**
- * @license React
- * react.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */
-
-/**
- * @license React
- * scheduler.production.min.js
- *
- * Copyright (c) Facebook, Inc. and its affiliates.
- *
- * This source code is licensed under the MIT license found in the
- * LICENSE file in the root directory of this source tree.
- */