Upload folder using huggingface_hub

Browse files

Files changed (5) hide show

h/cider-w8a8/README.md +84 -0
h/cider-w8a8/config.json +273 -0
h/cider-w8a8/manifest.json +39 -0
h/cider-w8a8/model.safetensors +3 -0
h/cider-w8a8/preprocessor_config.json +17 -0

h/cider-w8a8/README.md ADDED Viewed

	@@ -0,0 +1,84 @@

+---
+license: other
+license_name: nvidia-open-model-license
+base_model: nvidia/C-RADIOv4-H
+library_name: mlx
+pipeline_tag: image-feature-extraction
+tags:
+  - mlx
+  - c-radio
+  - vision
+  - embeddings
+  - quantized
+  - cider
+  - apple-silicon
+---
+# C-RADIOv4-H MLX Cider W8A8
+This is a self-contained MLX bundle converted from `nvidia/C-RADIOv4-H` for Apple
+Silicon image embeddings.
+Implementation repository:
+https://github.com/stephansturges/c-radio_v4_MLX
+## Source
+- Upstream model: https://huggingface.co/nvidia/C-RADIOv4-H
+- Upstream revision: `0057b339059c0b9e1b4ba996f975410ebbfdfcc8`
+- Converted bundle path in the local repo: `bundles/c-radiov4-h-cider-w8a8`
+## Format
+- Runtime: MLX plus Cider
+- Quantization: W8A8, per-channel int8 weights and online int8 activation quantization
+- Required hardware: Apple M5 or newer
+- Required package: https://github.com/Mininglamp-AI/cider
+- Quantized tensors: 129
+- Copied tensors: 261
+- Bundle size observed locally: 685 MB
+## Measured Accuracy
+Against the local bf16 MLX bundle on `data/golden_images/smoke.jpg`:
+| Image size | Summary cosine | Spatial cosine |
+| ---: | ---: | ---: |
+| 256x256 | 0.998008 | 0.997055 |
+| 512x512 | 0.997202 | 0.996210 |
+## Measured Speed
+Apple M5 Max, `mlx==0.31.2`, Cider `0.7.0`, no output materialization:
+| Resolution | Batch | p50 latency | Throughput |
+| ---: | ---: | ---: | ---: |
+| 256x256 | 1 | 15.5 ms | 64.6 images/s |
+| 256x256 | 4 | 37.4 ms | 106.9 images/s |
+| 512x512 | 1 | 47.1 ms | 21.2 images/s |
+| 512x512 | 4 | 179.6 ms | 22.3 images/s |
+## Usage
+Install Cider in a Python `>=3.12` environment, then run:
+```sh
+cradio-mlx embed \
+  --backend mlx-h \
+  --checkpoint /path/to/this/bundle \
+  --image image.jpg \
+  --image-size 512 \
+  --dtype bfloat16 \
+  --save-npz embedding.npz
+```
+## License
+The implementation code in `c-radio_v4_MLX` is MIT licensed. The model weights and this
+converted bundle are governed by NVIDIA's Open Model License Agreement, not by the MIT
+license. Preserve NVIDIA provenance and license terms when redistributing this bundle.
+NVIDIA Open Model License Agreement:
+https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf

h/cider-w8a8/config.json ADDED Viewed

	@@ -0,0 +1,273 @@

+{
+  "adaptor_configs": {},
+  "adaptor_names": null,
+  "architectures": [
+    "RADIOModel"
+  ],
+  "args": {
+    "aa": null,
+    "amp": true,
+    "amp_dtype": "bfloat16",
+    "amp_impl": "native",
+    "aug_repeats": 0,
+    "aug_splits": 0,
+    "auto_workload_inspector": false,
+    "bn_eps": null,
+    "bn_momentum": null,
+    "cache_dir": null,
+    "channels_last": false,
+    "checkpoint_folder": null,
+    "checkpoint_hist": 10,
+    "chk_keep_forever": 100,
+    "class_map": "",
+    "clip_grad": null,
+    "clip_mode": "norm",
+    "cls_token_per_teacher": true,
+    "coco_annotations_file": "/datasets/coco2017-adlsa/annotations/captions_val2017.json",
+    "coco_image_dir": "/datasets/coco2017-adlsa/val2017",
+    "color_jitter": 0.4,
+    "cooldown_epochs": 0,
+    "cpe_max_size": 2048,
+    "cpe_num_registers": null,
+    "crd_loss": false,
+    "crd_loss_weight": 0.8,
+    "crop_pct": null,
+    "cutmix": 0.0,
+    "cutmix_minmax": null,
+    "dataset_download": false,
+    "debug_full_knn": false,
+    "decay_epochs": 90,
+    "decay_milestones": [
+      90,
+      180,
+      270
+    ],
+    "decay_rate": 0.1,
+    "depchain": true,
+    "detect_anomaly": false,
+    "dist_bn": "reduce",
+    "dist_norm_weight": 0.0,
+    "distributed": true,
+    "drop": 0.0,
+    "drop_block": null,
+    "drop_connect": null,
+    "drop_path": null,
+    "dtype": "float32",
+    "epoch": 299,
+    "epoch_repeats": 0.0,
+    "eval": false,
+    "eval_metric": "knn_top1",
+    "eval_teacher": false,
+    "eval_teacher_only": false,
+    "eval_throughput": false,
+    "fast_norm": false,
+    "fd_loss_fn": "MSE",
+    "feature_normalization": "PHI_STANDARDIZE",
+    "feature_summarizer": "cls_token",
+    "feature_upscale_factor": null,
+    "force_disable_damp": false,
+    "force_disable_spectral_reparam": false,
+    "force_new_wandb_id": false,
+    "force_spectral_reparam": false,
+    "freeze_bn": false,
+    "fsdp": true,
+    "full_equivariance": false,
+    "fuser": "",
+    "gp": null,
+    "grad_accum_steps": 1,
+    "grad_checkpointing": false,
+    "head_init_bias": null,
+    "head_init_scale": null,
+    "head_lr": null,
+    "head_warmup": 3,
+    "head_weight_decay": 0.0005,
+    "hflip": 0.5,
+    "img_size": null,
+    "in_chans": null,
+    "initial_checkpoint": null,
+    "input_size": null,
+    "interpolation": "",
+    "layer_decay": null,
+    "local_rank": 0,
+    "log_interval": 50,
+    "log_mlflow": false,
+    "log_teacher_timings": true,
+    "log_train_metrics_per_epoch": true,
+    "log_train_metrics_per_log_interval": true,
+    "log_wandb": true,
+    "loss_auto_balance": false,
+    "lr_base": 0.1,
+    "lr_base_scale": "",
+    "lr_base_size": 256,
+    "lr_cycle_decay": 0.5,
+    "lr_cycle_limit": 1,
+    "lr_cycle_mul": 1.0,
+    "lr_k_decay": 1.0,
+    "lr_noise": null,
+    "lr_noise_pct": 0.67,
+    "lr_noise_std": 1.0,
+    "mean": null,
+    "mesa": false,
+    "min_lr": 1e-05,
+    "mixup": 0.0,
+    "mixup_mode": "batch",
+    "mixup_off_epoch": 0,
+    "mixup_prob": 1.0,
+    "mixup_switch_prob": 0.5,
+    "mlp_hidden_size": 1520,
+    "mlp_num_inner": 2,
+    "mlp_version": "v2",
+    "model": "vit_huge_patch16_224",
+    "model_kwargs": {},
+    "model_norm": false,
+    "momentum": 0.9,
+    "no_custom_validation": false,
+    "no_ddp_bb": true,
+    "no_knn": false,
+    "no_prefetcher": false,
+    "no_resume_opt": false,
+    "no_save_checkpoint": false,
+    "no_val": false,
+    "num_classes": null,
+    "on_demand_workload_inspector": false,
+    "one_logger_app_tag": "",
+    "one_logger_is_baseline": false,
+    "one_logger_run_name": "",
+    "onelogger": null,
+    "opt_betas": null,
+    "opt_eps": null,
+    "overfit": false,
+    "patience_epochs": 10,
+    "perf_test_no_aug": false,
+    "perf_test_no_decode": false,
+    "perf_test_no_io": false,
+    "perf_test_only_dataloader": false,
+    "perf_test_simple_aug": false,
+    "pin_mem": false,
+    "prefetcher": true,
+    "pretrained": false,
+    "processed_neck_outputs": null,
+    "profile_train_exit_after_profiling": false,
+    "profile_train_export_chrome_trace": true,
+    "profile_train_export_csv": false,
+    "profile_train_iterations": 0,
+    "qradio": false,
+    "qradio_max_tokens": 512,
+    "qradio_min_tokens": 32,
+    "qradio_patch_token_mask_initial_ratio": 0.95,
+    "qradio_progressive_2d": false,
+    "qradio_quantizer": null,
+    "qradio_ramp_alpha": 1.5,
+    "rank": 0,
+    "ratio": [
+      0.75,
+      1.3333333333333333
+    ],
+    "recount": 1,
+    "recovery_interval": 0,
+    "register_multiple": 10,
+    "remode": "pixel",
+    "reprob": 0.0,
+    "reset_loss_state": true,
+    "resplit": false,
+    "sample_tracking": false,
+    "save_images": false,
+    "scale": [
+      0.5,
+      1.0
+    ],
+    "sched": "cosine",
+    "seed": 42,
+    "shift_equivariance": false,
+    "smoothing": 0.1,
+    "source_tracking": false,
+    "spectral_heads": false,
+    "spectral_reparam": false,
+    "spectral_weight_decay": null,
+    "split_bn": false,
+    "start_epoch": null,
+    "std": null,
+    "stream_teachers": false,
+    "student_intermediate_indices": null,
+    "student_load_skip_state_dict_keys_regex": null,
+    "student_reinit_model_layers_regex": null,
+    "student_strict_load_ignore_mismatched_shape_keys_regex": null,
+    "student_strict_load_ignore_missing_keys_regex": null,
+    "student_strict_load_ignore_unexpected_keys_regex": null,
+    "student_strict_load_state_dict": false,
+    "sync_bn": false,
+    "sync_resolutions_across_ranks": true,
+    "synchronize_step": false,
+    "teachers": [
+      {
+        "model": "siglip2-g-384",
+        "name": "siglip2-g",
+        "spatial_mlp_version": "attn",
+        "type": "siglip2",
+        "use_summary": true
+      },
+      {
+        "model": "dinov3_vit7b16",
+        "name": "dino_v3_7b",
+        "type": "dino_v3",
+        "use_summary": true
+      },
+      {
+        "model": "default",
+        "name": "sam3",
+        "type": "sam3",
+        "use_summary": false
+      }
+    ],
+    "timing_warmup_iters": 20,
+    "tokenizer_kwargs": {},
+    "tokenizer_type": null,
+    "tome": null,
+    "torchcompile": null,
+    "torchscript": false,
+    "train_interpolation": "random",
+    "train_split": "train",
+    "tta": 0,
+    "untie_neck_weights": false,
+    "use_coco": false,
+    "use_multi_epochs_loader": false,
+    "val_ema_only": false,
+    "val_split": "val",
+    "vflip": 0.0,
+    "vitdet_version": 1,
+    "wandb_entity": "",
+    "wandb_id": "",
+    "wandb_job_type": "",
+    "wandb_name": "",
+    "wandb_project": "",
+    "wandb_tags": null,
+    "warmup_lr": 1e-05,
+    "warmup_prefix": false,
+    "worker_seeding": "all",
+    "workers": 8,
+    "workload_inspector_analyze_nsys_traces": false,
+    "workload_inspector_baseline_start_iter": 1500,
+    "workload_inspector_major_slowdown_p95_factor": 10.0,
+    "workload_inspector_minor_slowdown_p95_factor": 3.0,
+    "workload_inspector_no_slowdown_check": false,
+    "workload_inspector_simulate_slowdown_num_times": 1,
+    "workload_inspector_simulate_slowdown_start_iter": null,
+    "world_size": 256
+  },
+  "auto_map": {
+    "AutoConfig": "hf_model.RADIOConfig",
+    "AutoModel": "hf_model.RADIOModel"
+  },
+  "feature_normalizer_config": null,
+  "inter_feature_normalizer_config": null,
+  "max_resolution": 2048,
+  "patch_size": 16,
+  "preferred_resolution": [
+    512,
+    512
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.51.3",
+  "version": "c-radio_v4-h",
+  "vitdet_window_size": null
+}

h/cider-w8a8/manifest.json ADDED Viewed

	@@ -0,0 +1,39 @@

+{
+  "created_at": "2026-05-22T10:33:29Z",
+  "dtype": "bfloat16",
+  "extra": {
+    "conversion_state": "quantized_self_contained",
+    "quantization_stats": {
+      "copied_tensors": 261,
+      "padded_features": 0,
+      "padded_tensors": 0,
+      "quantized_tensors": 129
+    },
+    "source_bundle": "bundles/c-radiov4-h-bf16",
+    "weights_file": "model.safetensors"
+  },
+  "license": "nvidia-open-model-license",
+  "manifest_version": 1,
+  "max_resolution": 2048,
+  "model_id": "nvidia/C-RADIOv4-H",
+  "patch_size": 16,
+  "preferred_resolution": 512,
+  "quantization": {
+    "activation_bits": 8,
+    "bits": 8,
+    "group_size": 0,
+    "mode": "cider-w8a8",
+    "runtime": "cider",
+    "scheme": "symmetric_per_channel",
+    "state": "packed_w8a8_runtime",
+    "weight_bits": 8
+  },
+  "revision": "0057b339059c0b9e1b4ba996f975410ebbfdfcc8",
+  "source_files": {
+    "README.md": "be08bbaf49b38c62eb8496ebfd9b92292b4258838ade3d0e13e18d86978192d7",
+    "config.json": "f141520df919f1c29cd4f5a95340d9a39b15fbacafa36ca1054c55c919b2af62",
+    "model.safetensors": "f15d9d0c907974743a26a448da573daa2f2b75019fc1b9f31a45f0fdf49a7ee1",
+    "preprocessor_config.json": "6ca88d165592015b02a183d26a20fc96748e837d1cb21bde5f611141aca593cf"
+  },
+  "variant": "h"
+}

h/cider-w8a8/model.safetensors ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:f15d9d0c907974743a26a448da573daa2f2b75019fc1b9f31a45f0fdf49a7ee1
+size 717798440

h/cider-w8a8/preprocessor_config.json ADDED Viewed

	@@ -0,0 +1,17 @@

+{
+  "crop_size": {
+    "height": 512,
+    "width": 512
+  },
+  "do_center_crop": false,
+  "do_convert_rgb": true,
+  "do_normalize": false,
+  "do_rescale": true,
+  "do_resize": false,
+  "image_processor_type": "CLIPImageProcessor",
+  "processor_class": "CLIPProcessor",
+  "resample": 3,
+  "size": {
+    "shortest_edge": 512
+  }
+}