finalize pinned NVFP4A16 checkpoint metadata

README.md

@@ -39,12 +39,41 @@ This is **NVFP4A16**, not calibrated W4A4 NVFP4. Kimi K3 was trained
 with MXFP8 activations; FP4 activation quantization would require a separate
 calibration and quality-validation run.
 
-## Runtime status
+## Runtime validation (2026-07-27 UTC)
 
-**Runtime load has not been verified.** Static checkpoint structure and numeric
-conversion have been validated, but no end-to-end load or generation was run in
-the intended dedicated K3 vLLM/SGLang image. Do not treat structural validation
-as runtime compatibility or performance evidence.
+A purpose-built **vLLM one-layer routed-MoE smoke test passed** for this
+published artifact. The harness used vLLM's official compressed-tensors
+NVFP4A16 parameter layout, post-load conversion, fused-MoE method, and SiTU
+activation; it was not the full vLLM model loader or full K3 graph.
+
+- Evidence run: `20260727T220141Z`
+- Tested artifact revision: `8c22f32993e94e063cb634a03b7fc2f9ff539621`
+- Checkpoint shard: `model-00049-of-000096.safetensors` (17,916,197,160 bytes)
+- Runtime image: `vllm/vllm-openai:kimi-k3`
+- Pinned linux/amd64 image: `vllm/vllm-openai@sha256:fb16b180bd9727600067e16fcd6a6de43fb4db1baf4298ef20b4dbdf6bfa5a0e`
+- vLLM: `0.1.dev19262+gb6bbf29dd.d20260727`
+- Hardware: NVIDIA B200, SM100; torch `2.13.0+cu130`, CUDA 13.0
+- Exact MoE shape: 896 routed experts, top-k 16, `3584 -> 3072 -> 3584`, 2 tokens
+- Loaded tensors: 8,064 (`U8=2,688`, `F8_E4M3=2,688`, `F32=2,688`); all 896 experts
+- SiTU: beta `4.0`, linear beta `25.0`
+- Backend: `MARLIN` / `MarlinExperts`
+- Observed operators: `_moe_C::moe_wna16_marlin_gemm` (2 calls), Marlin CUDA
+  kernel, `_C::situ_and_mul`, `vllm::situ_and_mul_kernel`, routing alignment,
+  and MoE sum
+- Output: shape `[2, 3584]`, finite and nonzero
+
+**SGLang runtime was not executed.** A zero-cost gate against the official
+`lmsysorg/sglang:kimi-k3` image (linux/amd64 digest
+`sha256:2e8ef3746b2591287db0f37b4470910ab08974bdaf96fc5820bac35f6d3962bc`)
+found that its compressed-tensors selector supports FLOAT tensor-group NVFP4
+only as W4A4 with input activation scales; it has no matching NVFP4A16
+FLOAT tensor-group MoE path for `input_activations: null`. The gate therefore
+returned `BLOCKED / DO_NOT_RUN_GPU` (`20260727T221943Z`) before weight download
+or GPU allocation.
+
+Not verified: the full 93-layer / 2.8T model, KDA/MLA/vision integration,
+all 96 shards loaded together, end-to-end token generation, serving quality,
+or SGLang load/forward/kernel dispatch.
 
 ## License
 

nvfp4_conversion_manifest.json

@@ -18,5 +18,15 @@
   "scale_tensors": 247296,
   "tensor_bytes": 1645946964992,
   "lossless_over_span_policy": "fail; never silently requantize",
-  "runtime_load_verified": false
+  "runtime_validation": {
+    "level": "vllm_one_layer_routed_moe",
+    "vllm_verified": true,
+    "vllm_run_id": "20260727T220141Z",
+    "vllm_image_amd64_digest": "sha256:fb16b180bd9727600067e16fcd6a6de43fb4db1baf4298ef20b4dbdf6bfa5a0e",
+    "sglang_verified": false,
+    "sglang_static_gate": "BLOCKED",
+    "sglang_static_run_id": "20260727T221943Z",
+    "full_model_load_verified": false,
+    "full_generation_verified": false
+  }
 }