Spaces:
Sleeping
Sleeping
Commit History
-
feat: detect pre-quantized checkpoints, add NVFP4 presets, add test suite
- Fix a real correctness bug: HF's safetensors.total metadata reports the packed element count for bit-packed checkpoints (e.g. 2 values/byte for 4-bit formats), badly undercounting true parameter count. ModelAnalyzer now detects a quantization_config block (top-level, text_config, or modelopt-style config_groups — scanning all groups and taking the narrowest format, since some checkpoints mix precisions per-layer) and falls back to the architecture-derived total instead. Generalizes to any HF repo typed in directly, not just curated presets. - Add an explicit total_params override field for presets where the exact count is already known from elsewhere (e.g. a bf16 twin's safetensors). - Fix a second bug hit while testing the above: dynamically-fetched MoE configs often split a much smaller per-expert moe_intermediate_size from an unrelated (sometimes null) base intermediate_size field. Reading it naively inflated a real ~36B-parameter model's estimate to 516B. Fi…4e1e495
-
feat: add multimodal models, TTFT-driven GPU sizing, and fix param/KV bugs
- Add Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B, and google/gemma-4-26B-A4B-it (the MoE Gemma 4 variant) — all natively text+image models, verified against their real HF config.json. - Fix head_dim: it was always derived as hidden_size // num_attention_heads, which is wrong for Gemma, gpt-oss, and the whole Qwen3.5/3.6 family (they use a fixed head_dim independent of that ratio). Now read explicitly from config when present. - Fix Qwen3-VL-235B-A22B-* and Qwen3-VL-30B-A3B-* presets: hidden_size, layer count, kv-heads, and expert count/width didn't match the real configs (30B-A3B was missing MoE data entirely and was sized as dense). - Fix notes.moe fallback only recognizing num_local_experts, silently defaulting to 8 experts for entries using num_experts. - Replace the flat 80%-of-params-in-experts heuristic with an architecture-derived expert/non-expert split applied to the real (safetensors) total — active-param estimates for sparse MoE models were off by up to 2.7…0f64354
-
fix: correct B200 server recommendations to SR685a V3 / SR680a V3
SR675 V3 does not support the HGX B200 (SXM6 form factor). Valid options per Lenovo's HGX B200 product guide are SR685a V3 and SR680a V3.ddbbeb1
-
fix: apply overhead to dynamic memory only, fix KV bytes and ITL bandwidth
- Overhead % was incorrectly applied to the full memory footprint including static weights, inflating requirements and causing a 3rd GPU to appear for 70B models that fit in 2. Now applied to dynamic memory only, plus a 1 GB fixed CUDA context constant. - KV cache byte width was hardcoded to 2 regardless of quantization. Now uses bytes_per_param so INT8 and FP4 paths size correctly. - ITL bandwidth calculation was using per-user KV instead of the full concurrent KV footprint; decode must stream all active KV caches each step. - Expand README with full explanation of the sizing model, latency formulas, config file schemas, and a glossary of all acronyms.33f22e2
-
feat: add Qwen3.5 models and NVIDIA RTX PRO 6000 Blackwell GPUs
- Add 8 Qwen3.5 model presets (0.8B–397B, dense and MoE variants) - Add RTX PRO 6000 Blackwell Workstation (96GB, 1792 GB/s) and Server (96GB, 1597 GB/s) editions Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>2271b6a
-
feat: add RAG support, improve parameter calculation, and enhance UI
- Add RAG pipeline support with embedding and reranker model selection (WIP) - Improve parameter calculation using HuggingFace API safetensors metadata - Enhance MoE detection and active parameter calculation - Add configurable GPU memory overhead percentage slider - Improve network/interconnect logic with PCIe bottleneck handling - Add Qwen3-VL models (235B and 30B variants) to models.yaml - Reorder hardware_data.yaml by cost_tier for better UX - Remove legend from donut chart (hover tooltips provide info) - Update memory breakdown to include RAG models category - Fix text_config handling for vision-language models1968f0b -
feat: convert memory chart to interactive donut chart and improve descriptions
- Replace matplotlib pie chart with plotly donut chart for better interactivity - Add hover tooltips showing detailed memory breakdown per segment - Update color palette to modern, accessible colors - Enhance memory breakdown text description with category explanations - Add plotly>=5.0.0 to dependencies (pyproject.toml and requirements.txt) - Fix Gradio 6.0 compatibility: move theme and css to launch() method - Update default slider values: 50 users, 1024 input context, 256 output tokens - Improve WCAG compliance with better text alternatives and descriptions83ef6b4
-
Initial commit
3846650 -
initial commit
Verified signature3d49912