Update model card with inter-step cache benchmark results (v0.2.0)

README.md

@@ -30,22 +30,21 @@ LLaDA is a **diffusion language model** that generates text by iterative unmaski
 
 ## Benchmark (AMD EPYC 4465P 12-Core, steps=16, threads=12)
 
-### Real Prompt Performance (Q4_K_M + entropy_exit)
+### Real Prompt Performance (Q4_K_M + entropy_exit + inter-step cache, B=256)
 
-| Prompt | B=64 tok/s | B=256 tok/s | Steps | vs llama.cpp |
+| Prompt | No-Cache tok/s | Cache tok/s | Steps | vs llama.cpp |
 |---|---|---|---|---|
-| Capital of France? | 9.22 | **15.60** | 4 | 1.8x |
-| Translate to French | 10.23 | **21.78** | 3 | 2.6x |
-| 15 × 23? | 11.49 | **11.45** | 5 | 1.3x |
-| Translate to Spanish | 4.59 | **7.17** | 8 | 0.8x |
-| Python is_prime() | 2.53 | **3.12** | 17 | 0.4x |
-| Poem about ocean | 2.33 | **3.10** | 17 | 0.4x |
-| Why is sky blue? | 2.21 | **3.18** | 17 | 0.4x |
-| List the planets | 2.33 | **3.19** | 17 | 0.4x |
-
-*B = generation buffer size (tokens generated per call). llama.cpp baseline: 8.51 tok/s (Llama-3-8B Q4_K_M, same hardware).*
-
-entropy_exit adapts to prompt difficulty: 3–4 steps for easy, 16 for hard. Never slower than baseline.
+| Capital of France? | 17.5 | **24.4** | 3 | 2.9x |
+| Translate to French | 25.9 | **27.7** | 2 | 3.3x |
+| 15 x 23? | 12.8 | **15.7** | 4 | 1.8x |
+| Translate to Spanish | 7.6 | **22.9** | 7 | 2.7x |
+| Python is_prime() | 3.2 | **4.9** | 16 | 0.6x |
+| Poem about ocean | 3.2 | **5.3** | 16 | 0.6x |
+| Why is sky blue? | 3.3 | **12.0** | 16 | 1.4x |
+| List the planets | 3.3 | **9.4** | 15 | 1.1x |
+| **Average** | **9.6** | **15.3** | | **1.8x** |
+
+*llama.cpp baseline: 8.51 tok/s (Llama-3-8B Q4_K_M, same hardware). Cache enabled by default. 6 of 8 prompts outperform llama.cpp; 2 (code generation, creative writing) remain slower due to requiring all 16 steps.*
 
 ### Quantization Comparison (low_confidence baseline, B=64)
 
@@ -57,8 +56,10 @@ entropy_exit adapts to prompt difficulty: 3–4 steps for easy, 16 for hard. Nev
 
 ### Summary
 
-- **11–22 tok/s on easy real prompts** (Q4_K_M + entropy_exit, B=256)
-- **Up to 2.6x faster than llama.cpp** on the same hardware
+- **15-28 tok/s on easy real prompts** (Q4_K_M + entropy_exit + inter-step cache, B=256)
+- **Up to 3.2x faster than llama.cpp** on the same hardware
+- **Inter-step KV cache**: 1.6x average speedup with no quality degradation
+- **6 of 8 real prompts outperform llama.cpp** (vs 3 of 8 without cache)
 - **256-token generation** with 20% lower per-token cost vs 64-token batches
 - **7.5x thread scaling** from 1 to 12 threads
 
@@ -72,7 +73,7 @@ cd diffuse-cpp
 cmake -B build -DCMAKE_BUILD_TYPE=Release
 cmake --build build -j$(nproc)
 
-# Generate with entropy_exit (recommended)
+# Generate with entropy_exit + cache (recommended, cache is ON by default)
 python tools/generate.py \
     --model-dir /path/to/LLaDA-8B-Instruct \
     --gguf llada-8b-q4km.gguf \