Publish qwen-3-5-0-8b-q4k-ehaf16 registry metadata

registry/catalog.json

@@ -1,7 +1,7 @@
 {
   "version": 1,
   "lifecycleSchemaVersion": 1,
-  "updatedAt": "2026-04-18",
+  "updatedAt": "2026-04-20",
   "models": [
     {
       "modelId": "gemma-3-270m-it-q4k-ehf16-af32",
@@ -226,12 +226,13 @@
         "tested": {
           "suite": "inference",
           "surface": [
+            "browser",
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-18",
-          "source": "runtime-verify",
-          "notes": "Apple M3 / Node WebGPU darwin re-verified 2026-04-18 against the current local manifest; execution contract pass. Receipt: reports/gemma-4-e2b-it-q4k-ehf16-af32/2026-04-18T19-17-45.997Z.json. Previous AMD RDNA-3 re-verify 2026-04-09 after repinning Gemma 4 E2B text decode to attention_decode_online_f16kv and splitting prefill attention between sliding-window and full-attention layers. Apple M3 browser compare receipt: benchmarks/vendors/results/compare_20260415T170108.json (Doppler RDRR wins cold load ~3.4x, Transformers.js ONNX/q4f16 wins warm decode ~2.6x; product-engine comparison across different artifact formats, not a format-identical kernel benchmark). Manifest-only republish to HF 2026-04-18 at revision c690ac0274891afbddb8afb776d000e4c0051b7b.",
+          "lastVerifiedAt": "2026-04-20",
+          "source": "vendor-compare",
+          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-20 with release decode cadence. Compare receipts benchmarks/vendors/results/compare_20260420T162522.json (t=0 greedy) and benchmarks/vendors/results/compare_20260420T162831.json (t=1 topK50) show Doppler decode throughput ahead of the paired Transformers.js ONNX/q4f16 runner on warm OPFS. Apple M3 / Node WebGPU darwin execution contract pass remains from reports/gemma-4-e2b-it-q4k-ehf16-af32/2026-04-18T19-17-45.997Z.json. Manifest-only republish to HF 2026-04-18 at revision c690ac0274891afbddb8afb776d000e4c0051b7b.",
           "contracts": {
             "executionContractOk": true
           }
@@ -308,7 +309,7 @@
       "modelId": "qwen-3-5-0-8b-q4k-ehaf16",
       "family": "qwen3",
       "label": "Qwen 3.5 0.8B (Q4K)",
-      "description": "Qwen 3.5 0.8B text model with hybrid linear attention. Q4K weights, F16 embeddings/head/accumulation.",
+      "description": "Qwen 3.5 0.8B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
       "modes": [
         "text",
         "vision"
@@ -324,7 +325,7 @@
       },
       "hf": {
         "repoId": "Clocksmith/rdrr",
-        "revision": "1c509e37e4430d7ffd094ce5dfa623792603670e",
+        "revision": "4968d761a8135ddace5bf98cc0423be28ce0b3db",
         "path": "models/qwen-3-5-0-8b-q4k-ehaf16"
       },
       "baseUrl": null,
@@ -347,12 +348,13 @@
         "tested": {
           "suite": "inference",
           "surface": [
+            "browser",
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-18",
-          "source": "runtime-verify",
-          "notes": "Re-verified 2026-04-18 on AMD Strix Halo / Radeon 8060S (RDNA-3) + RADV Mesa 26.0.3, Node/WebGPU after kernel-ref swap on the 6 full-attention layers (indices 3/7/11/15/19/23): attn_stream -> attn_head256, and q4_prefill multicol_shared -> q4_widetile (register-tiled). Warm-cache 3-run means showed prefill@80 +57.5% (sigma=3.8), prefill@15 +16.6%, decode flat (sigma=1.3); doppler verify match with the expected Qwen <think> wrapper; execution contract gate pass. Receipts live on the Strix Halo box (this darwin workspace only carries Apr 15 pre-swap receipts under reports/qwen-3-5-0-8b-q4k-ehaf16/). Previous 2026-04-17 run confirmed the Apr-17 profiles/qwen-3-5-0-8b-throughput retune (batchSize=4 readbackInterval=2 batch stop-check); 2026-04-13 Apple M3 browser verification still applies to that surface. Manifest-only republish to HF 2026-04-18 at revision 1c509e37e4430d7ffd094ce5dfa623792603670e.",
+          "lastVerifiedAt": "2026-04-20",
+          "source": "vendor-compare",
+          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-20 after the quantized tied-LM-head update. Compare receipts benchmarks/vendors/results/compare_20260420T163342.json (t=0 greedy) and benchmarks/vendors/results/compare_20260420T163425.json (t=1 topK50) show Doppler decode throughput ahead of the paired Transformers.js runner on warm OPFS. Earlier 2026-04-18 Node/WebGPU verification after the Qwen prefill kernel-ref swap remains valid; execution contract gate pass. Manifest-only republish to HF 2026-04-18 at revision 1c509e37e4430d7ffd094ce5dfa623792603670e.",
           "contracts": {
             "executionContractOk": true
           }
@@ -369,7 +371,7 @@
       "modelId": "qwen-3-5-2b-q4k-ehaf16",
       "family": "qwen3",
       "label": "Qwen 3.5 2B (Q4K)",
-      "description": "Qwen 3.5 2B text model with hybrid linear attention. Q4K weights, F16 embeddings/head/accumulation.",
+      "description": "Qwen 3.5 2B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
       "modes": [
         "text",
         "vision"
@@ -412,9 +414,9 @@
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-18",
-          "source": "runtime-verify",
-          "notes": "Apple M3 / Node WebGPU darwin re-verified 2026-04-18 against the current local manifest; execution contract pass. Receipt: reports/qwen-3-5-2b-q4k-ehaf16/2026-04-18T19-18-27.036Z.json. Node/WebGPU verification uses execution-v1 with fused-Q4 decode/prefill projections as the primary hybrid Qwen path and execution.inlineKernelPath=true. The earlier GEMV/head256 note was rolled back after correctness regressions on the linear-attention path. Manifest-only republish to HF 2026-04-18 at revision 3f54bc54c19181bd737d1773dcd70fb783297e73.",
+          "lastVerifiedAt": "2026-04-20",
+          "source": "vendor-compare",
+          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-20 after the Qwen 3.5 2B retune; execution contract pass with manifest-owned decodeBatchSize=12. Receipt: reports/qwen-3-5-2b-q4k-ehaf16/2026-04-20T12-49-30.650Z.json. Current execution-v1 uses the fixed fused-Q4 main_gemv transformer decode path, keeps stable fused-Q4 main_multicol declared as fallback, upgrades the tied LM head to Q4 via optimized lm_head_q4 GEMV 64x4, upgrades prefill projections to q4_widetile, and upgrades prefill attention to attn_head256. Compare receipts benchmarks/vendors/results/compare_20260420T163143.json (t=0 greedy) and benchmarks/vendors/results/compare_20260420T163255.json (t=1 topK50, run with --allow-non-comparable-lane because the lane remains capability_only) show Doppler decode throughput ahead of the paired Transformers.js runner on warm OPFS. Manifest-only republish to HF 2026-04-18 at revision 3f54bc54c19181bd737d1773dcd70fb783297e73.",
           "contracts": {
             "executionContractOk": true
           }