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

Dockerfile

@@ -364,12 +364,13 @@ 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 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=83333 \
+    # Validation limits are per deepsearch-api process: 32 checks in flight and a 31k-verdict cache in
+    # each of the 16 processes (DEEPSEARCH_API_PROCESSES below), 512 and ~500k in all (Redis keeps every verdict
+    # anyway). One search's page runs at most 16 at once (usdsearch's BATCH_CONCURRENCY). 2026-10-20 lab, l4x4
+    # shape with the check-image cache: 16 per process capped the engine side at ~155 checks/s (p95 of all badges
+    # 14 s at 80 users); 32 held 3.2 s up to 240 users.
+    USDSEARCH_VISION_VALIDATION_MAX_CONCURRENT=32 \
+    USDSEARCH_VISION_VALIDATION_CACHE_SIZE=31250 \
     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.
@@ -400,13 +401,14 @@ ENV HF_STORAGE_API_TOKEN= \
     TRITON_SERVER_URL=localhost:8001 \
     SIGLIP2_TRITON_SERVER_SSL=false \
     # 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
+    # upstream (supervisord-gpu.conf, nginx.conf.template). 16 for l4x4 (48 vCPU):
+    # 2026-10-20 lab, l4x4 shape, 32 checks each: 16 processes held p95 of all badges at
+    # 3.2 s up to 240 users (6.3 s at 300, CPU-bound); 6 processes reached ~100 users
+    # with the CPU under 30 %. On a10g-large (12 vCPU) use 6: 8 cost ~20 % more CPU there. 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 \
+    DEEPSEARCH_API_PROCESSES=16 \
     PORT=7860
 
 ARG RUN_ID

README.md

@@ -6,7 +6,7 @@ colorTo: blue
 sdk: docker
 app_port: 7860
 startup_duration_timeout: 1h
-hardware: a10g-large
+hardware: l4x4
 ---
 
 # SimReady Central USD Search
@@ -14,7 +14,7 @@ hardware: a10g-large
 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 (a10g-large), unlike the free-tier
+hardware tier (l4x4), 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