Sync model repo (text/metadata)

README.md

@@ -29,9 +29,9 @@ This version is intended to demonstrate efficient inference on Arm-based platfor
 | Model format | ONNX |
 | Target device class | Mobile CPU |
 | Reference device | vivo X300 (C1-Ultra, C1-Premium, C1-Pro, Android 16 / OriginOS 6) |
-| Primary performance result | 31.88 tokens/sec decode throughput (3.83x the FP32 baseline); 1175.59 ms time to first token, which is 25.2 % slower than the FP32 baseline |
+| Primary performance result | 34.80 tokens/sec decode throughput (4.27x the FP32 baseline); 726.50 ms time to first token, which is 22.6 % faster than the FP32 baseline |
 | Accuracy result | HellaSwag 59.0 % |
-| Size / memory result | 727.86 MB with INT4 weights, INT8 activations, INT8 embeddings and INT8 LM head (5.77x smaller than the FP32 baseline); peak memory 879.30 MB |
+| Size / memory result | 727.86 MB with INT4 weights, INT8 activations, INT8 embeddings and INT8 LM head (5.77x smaller than the FP32 baseline); peak memory 897.54 MB |
 
 ## Original model
 
@@ -63,27 +63,29 @@ Performance was measured on the reference configuration below. Results are inten
 | Field | Value |
 |---|---|
 | Device / platform | vivo X300 |
-| CPU / accelerator | C1-Ultra, C1-Premium, C1-Pro, aarch64, 8 cores, CPU execution backend |
+| CPU / accelerator | C1-Ultra, C1-Premium, C1-Pro, aarch64, CPU execution backend |
 | OS | android, Android 16 / OriginOS 6 |
 | Runtime | ONNX Runtime 1.27.0 |
 | Backend / delegate | MLAS, KleidiAI |
 | Batch size | 1 |
 | Precision | INT4 groupwise asymmetric weights (GPTQ, group size 32), INT8 per-token dynamic activations, INT8 per-row embeddings, INT8 LM head (scheme W4A8_dyn_emb_int8) |
-| Runs | 20 warmup runs + 100 measured runs |
+| Runs | 5 warmup runs + 20 measured runs |
+
+**Measurement conditions.** Each measured run consumes a 153-token prompt and generates 128 tokens, using 4 CPU threads, 5 warm-up runs and 20 measured runs. Each run starts only once Android reports thermal status 0 (NONE), after a 30 s settle. The device is set to fixed performance mode, which is the official recommendation.
 
 **Performance results**
 
 | Metric | Original / baseline | Arm-optimized | Improvement |
-|---|---:|---:|---:|
-| End-to-end latency p50 (ms) | 4549.87 | 2085.09 | 2.18x faster |
-| End-to-end latency p90 (ms) | 4700.35 | 2223.47 | 2.11x faster |
-| End-to-end latency p99 (ms) | 4818.38 | 2354.50 | 2.05x faster |
-| Decode throughput (tokens/sec) | 8.31 | 31.88 | 3.83x |
-| Time to first token (ms) | 938.81 | 1175.59 | 0.80x (regression) |
-| Model load time (ms) | 9086.10 | 1622.18 | 5.60x faster |
+| --- | ---: | ---: | ---: |
+| End-to-end latency p50 (ms) | 16658.09 | 4394.52 | 3.79x faster |
+| End-to-end latency p90 (ms) | 16835.99 | 4727.38 | 3.56x faster |
+| End-to-end latency p99 (ms) | 16910.87 | 4769.23 | 3.55x faster |
+| Decode throughput (tokens/sec) | 8.15 | 34.80 | 4.27x |
+| Time to first token (ms) | 938.04 | 726.50 | 1.29x faster |
+| Model load time (ms) | 7604.93 | 1358.55 | 5.60x faster |
 | Model size (MB) | 4196.99 | 727.86 | 5.77x smaller |
-| Peak memory (MB) | 4073.95 | 879.30 | 4.63x less |
-| Average memory (MB) | 4054.65 | 860.36 | 4.71x less |
+| Peak memory (MB) | 4098.56 | 897.54 | 4.57x less |
+| Average memory (MB) | 4079.43 | 877.93 | 4.65x less |
 
 ## Accuracy
 

benchmarks/tinyllama-1-1b-chat-onnx-genai-vivo-x300-fp32.yaml

@@ -1,7 +1,7 @@
 version: 1.2.0
 report_type: llm-generative
 profile: Baseline
-created_at: '2026-07-10T13:41:20Z'
+created_at: '2026-09-04T20:59:14Z'
 context:
   target_ref: vivo-x300
   runtime:
@@ -20,22 +20,22 @@ context:
     sample_count: 8000
   benchmark:
     batch_size: 1
-    num_runs: 100
-    warmup_runs: 20
-    prompt_length_tokens: 154
-    generation_length_tokens: 30
+    num_runs: 20
+    warmup_runs: 5
+    prompt_length_tokens: 153
+    generation_length_tokens: 128
 accuracy:
   benchmark_name: HellaSwag
   accuracy_pct: 59.8
   shot_count: 0
 performance:
   end_to_end_latency_ms:
-    p50: 4549.870
-    p90: 4700.350
-    p99: 4818.378
-  peak_memory_mb: 4073.949
-  average_memory_mb: 4054.648
-  ttft_ms: 938.808
-  tokens_per_second: 8.315
-  model_load_time_ms: 9086.103
-  time_to_first_inference_ms: 4535.913
+    p50: 16658.093
+    p90: 16835.987
+    p99: 16910.870
+  peak_memory_mb: 4098.555
+  average_memory_mb: 4079.430
+  ttft_ms: 938.042
+  tokens_per_second: 8.148
+  model_load_time_ms: 7604.926
+  time_to_first_inference_ms: 16420.157

benchmarks/tinyllama-1-1b-chat-onnx-genai-vivo-x300-int4.yaml

@@ -1,7 +1,7 @@
 version: 1.2.0
 report_type: llm-generative
 profile: Arm-Optimized
-created_at: '2026-07-09T02:55:51Z'
+created_at: '2026-09-04T20:59:14Z'
 context:
   target_ref: vivo-x300
   runtime:
@@ -20,23 +20,23 @@ context:
     sample_count: 8000
   benchmark:
     batch_size: 1
-    num_runs: 100
-    warmup_runs: 20
-    prompt_length_tokens: 154
-    generation_length_tokens: 30
+    num_runs: 20
+    warmup_runs: 5
+    prompt_length_tokens: 153
+    generation_length_tokens: 128
 accuracy:
   benchmark_name: HellaSwag
   accuracy_pct: 59.0
   shot_count: 0
 performance:
   end_to_end_latency_ms:
-    p50: 2085.095
-    p90: 2223.471
-    p99: 2354.495
-  peak_memory_mb: 879.301
-  average_memory_mb: 860.359
+    p50: 4394.521
+    p90: 4727.384
+    p99: 4769.225
+  peak_memory_mb: 897.539
+  average_memory_mb: 877.934
   delegation_pct: 99.84
-  ttft_ms: 1175.594
-  tokens_per_second: 31.881
-  model_load_time_ms: 1622.184
-  time_to_first_inference_ms: 2091.532
+  ttft_ms: 726.502
+  tokens_per_second: 34.800
+  model_load_time_ms: 1358.552
+  time_to_first_inference_ms: 4482.815