Upload README.md with huggingface_hub

README.md

@@ -119,6 +119,34 @@ The first digit after the decimal is 1 for 9.11 and 9 for 9.90. Since 9 > 1,
 So, 9.9 is bigger than 9.11.
 ```
 
+## Performance (measured)
+
+Test environment — the resources **actually usable inside the container (docker)**:
+
+| Resource | Value |
+|---|---|
+| CPU | Intel Xeon Platinum 8559C |
+| Effective cores | **16 vCPU** (cgroup `cpu.max` = `1600000 100000`) |
+| Memory (cgroup limit) | **≈ 97 GiB** |
+| GPU | **none usable** — `/dev/nvidia*` nodes exist but `cuInit()` returns `CUDA_ERROR_NO_DEVICE` and NVML init fails |
+| Build / runtime | CPU-only, `-t 16` |
+
+Measured on `pangu-diffusion-Q4_K_M.gguf` (4.59 GiB, 8.03 B params):
+
+| Mode | Benchmark | Speed |
+|---|---|---|
+| **Input** — prompt processing (causal prefill) | `llama-bench` pp512 | **131.8 t/s** (131.77 ± 3.34) |
+| **Output** — autoregressive decoding | `llama-bench` tg128 | **11.1 t/s** (11.11 ± 0.35) |
+| **Output** — block diffusion (block 32 / small-block 4, `alg=entropy`, `temp=0`) | 224 tokens | **≈ 4.4 t/s** (≈ 55 s end-to-end) |
+
+Notes:
+- The input/prompt-processing number is the same causal prefill used by both the AR and the
+  diffusion path.
+- Block diffusion reveals one token per denoising step, i.e. it runs up to `block_length`
+  (= 32) extra forward passes per block, so its tokens/s is inherently lower than AR.
+- The AR figures (`pp512` / `tg128`) are the best reference for raw model throughput on
+  this hardware.
+
 ## What the patch does
 
 1. **`src/llama-arch.cpp`** — mark `LLM_ARCH_PANGU_EMBED` as a diffusion architecture