From 6333be7b83d97fa609df61b650032c49a71d5867 Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Sat, 5 Sep 2026 12:42:01 +0200 Subject: [PATCH] cleanups Signed-off-by: Vladimir Mandic --- cli/sdnq-attention-benchmark.py | 4 ++-- data/reference-cloud.json | 12 ++++++++---- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/cli/sdnq-attention-benchmark.py b/cli/sdnq-attention-benchmark.py index 7e3ddf354..fbcf9fb36 100755 --- a/cli/sdnq-attention-benchmark.py +++ b/cli/sdnq-attention-benchmark.py @@ -465,7 +465,7 @@ def load_sdnext(): raise detail = f"exited with code {e.code}" if isinstance(e, SystemExit) else f"{type(e).__name__}: {e}" console.print(f"[red]sdnext failed to start: {detail}[/red]") - text = startup_log["text"].strip() + text = startup_log["text"].strip() # pylint: disable=used-before-assignment if text: console.print(Panel(escape(text[-4000:]), title="sdnext startup log", box=ROUNDED_BOX)) console.print("run from the sdnext root with the venv active; triton is required") @@ -2229,7 +2229,7 @@ def resolved_group_label(layer, in_features): return "row" -def bench_group_sizes(shape_label, out_features, in_features, plain_results, selected_dtypes, iters, warmup, config_timeout=300): +def bench_group_sizes(shape_label, out_features, in_features, plain_results, selected_dtypes, iters, warmup, config_timeout=300): # pylint: disable=unused-argument # the Group size setting: 0 = auto, -1 = row-wise, explicit values snap to a divisor of # in_features; grouping forces a per-forward re-quantize when quantized matmul is on, so # the mm cells price that cost alongside the accuracy gain diff --git a/data/reference-cloud.json b/data/reference-cloud.json index 10f0e5930..59327c848 100644 --- a/data/reference-cloud.json +++ b/data/reference-cloud.json @@ -2,21 +2,25 @@ "Nano Banana lite": { "path": "gemini-3.1-flash-lite-image", "desc": "Nano Banana Lite is designed as the efficiency specialist of the image generation family, offering ultra-low latency and cost-effective image generation and editing. By targeting a sub-2 second latency and significantly reduced TPU compute costs, this model enables high-volume interactive developer use cases and real-time consumer applications. (Knowledge cutoff January 2025)", - "preview": "gemini-3.1-flash-lite-image.jpg" + "preview": "gemini-3.1-flash-lite-image.jpg", + "date": "2026 July" }, "Nano Banana 2": { "path": "gemini-3.1-flash-image", "desc": "Nano Banana 2 provides high-quality image generation and conversational editing at a mainstream price point and low latency. It serves as the high-efficiency counterpart to Gemini 3 Pro Image, optimized for speed and high-volume developer use cases.(Knowledge cutoff January 2025)", - "preview": "gemini-3.1-flash-image.jpg" + "preview": "gemini-3.1-flash-image.jpg", + "date": "2026 February" }, "Nano Banana Pro": { "path": "gemini-3-pro-image", "desc": "Nano Banana Pro is a sophisticated reasoning-driven engine for professional-grade image editing and generation, offering studio-quality precision and advanced creative control. Nano Banana Pro is best for complex graphic design, high-fidelity product mockups, and factual data visualizations that require accurate text rendering and real-world grounding via Google Search. (Knowledge cutoff January 2025)", - "preview": "gemini-3-pro-image.jpg" + "preview": "gemini-3-pro-image.jpg", + "date": "2025 November" }, "Nano Banana": { "path": "gemini-2.5-flash-image", "desc": "Our best engine for high-velocity visual creation, offering state-of-the-art speed and efficiency. Gemini 2.5 Flash Image, also known as Nano Banana, is best for high-volume generation, conversational image editing, and low-latency creative workflows that require native multimodal understanding. (Knowledge cutoff June 2025)", - "preview": "gemini-2.5-flash-image.jpg" + "preview": "gemini-2.5-flash-image.jpg", + "date": "2025 August" } }