Add torch_cuda deployment: checkpoints + updated manifest (ascend_310p preserved)

assets/adapter.json

@@ -1,9 +1,8 @@
 {
   "diffusion_steps": 10,
-  "family": "graspgen",
+  "discriminator_checkpoint": "assets/discriminator_checkpoint.pth",
+  "generator_checkpoint": "assets/generator_checkpoint.pth",
   "geometry": {
-    "ball_query_order": "input_index",
-    "fps_start_index": 0,
     "npoints": [
       256,
       64
@@ -18,8 +17,18 @@
     ]
   },
   "grasp_batch_size": 1000,
+  "gripper_config": "assets/graspgen_config.yml",
+  "interface": "tensor_model",
   "kappa": 2.02217,
+  "model_type": "graspgen",
+  "operation": "generate_grasps",
   "point_count": 2048,
   "postprocessing": "grasp-pose-matrix-confidence-sorted-v1",
-  "preprocessing": "object-points-outlier-filtered-centered-kappa-scaled-v1"
+  "preprocessing": "object-points-outlier-filtered-centered-kappa-scaled-v1",
+  "source_sha256": {
+    "assets/discriminator_checkpoint.pth": "52f68d3cd3decb71499e084fdb88808d174419a72cd9ef38f821083a4409895d",
+    "assets/generator_checkpoint.pth": "fe8497108e39d8fc50be06cd7df22a2f680e0495e713d24c3104616f291e00dd",
+    "assets/graspgen_config.yml": "93804f423ec02def08add164d83379212204ddc33779cd273ffffdacf94f67f3"
+  },
+  "torch_module_loader": "perception_service.torch_model_loaders:load_graspgen"
 }

assets/discriminator_checkpoint.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52f68d3cd3decb71499e084fdb88808d174419a72cd9ef38f821083a4409895d
+size 39427928

assets/generator_checkpoint.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe8497108e39d8fc50be06cd7df22a2f680e0495e713d24c3104616f291e00dd
+size 471436414

assets/graspgen_config.yml

@@ -0,0 +1,245 @@
+data:
+  root_dir: ''
+  tasks:
+  - pick
+  num_points: 2048
+  num_object_points: 1024
+  cam_coord: false
+  num_rotations: 64
+  grid_resolution: 0.01
+  jitter_scale: 0.0
+  contact_radius: 0.005
+  dist_above_table: 0.002
+  robot_prob: 1.0
+  offset_bins:
+  - 0
+  - 0.00794435329
+  - 0.0158887021
+  - 0.0238330509
+  - 0.0317773996
+  - 0.0397217484
+  - 0.0476660972
+  - 0.055610446
+  - 0.0635547948
+  - 0.0714991435
+  - 0.08
+  random_seed: -1
+  rotation_augmentation: false
+  add_depth_noise: false
+  downsample_points: false
+  downsample_to_fixed_batch_size: false
+  dataset_cls: ObjectPickDataset
+  load_patch: false
+  load_discriminator_dataset: false
+  patch_width: 256
+  prob_point_cloud: -1
+  object_root_dir: ''
+  grasp_root_dir: ''
+  dataset_name: acronym
+  dataset_version: v1
+  gripper_name: robotiq_2f_140
+  num_grasps_per_object: 1000
+  load_contact: false
+  prefiltering: false
+  discriminator_ratio:
+  - 0.5
+  - 0.2
+  - 0.25
+  - 0.05
+  - 0.0
+  visualize_batch: false
+  onpolicy_dataset_dir: null
+  onpolicy_dataset_h5_path: null
+  redundancy: 1
+  preload_dataset: false
+m2t2:
+  scene_encoder:
+    type: pointnet2_msg
+    num_points: 16384
+    downsample: 4
+    radius: 0.02
+    radius_mult: 2
+    use_rgb: false
+  object_encoder:
+    type: pointnet2_msg_cls
+    num_points: 1024
+    downsample: 4
+    radius: 0.02
+    radius_mult: 2
+    use_rgb: false
+  contact_decoder:
+    mask_feature: res0
+    in_features:
+    - res1
+    - res2
+    - res3
+    place_feature: res4
+    embed_dim: 256
+    feedforward_dim: 512
+    num_scales: 3
+    num_layers: 3
+    num_heads: 8
+    num_grasp_queries: 1000
+    num_place_queries: 0
+    use_attn_mask: true
+    use_task_embed: false
+    activation: GELU
+    pos_enc: new
+  action_decoder:
+    use_embed: false
+    conf_thresh: 0.5
+    max_num_pred: null
+    hidden_dim: 256
+    num_layers: 2
+    activation: GELU
+    offset_bins:
+    - 0
+    - 0.00794435329
+    - 0.0158887021
+    - 0.0238330509
+    - 0.0317773996
+    - 0.0397217484
+    - 0.0476660972
+    - 0.055610446
+    - 0.0635547948
+    - 0.0714991435
+    - 0.08
+    gripper_depth: 0.1034
+    gripper_name: robotiq_2f_140
+  matcher:
+    object_weight: 2.0
+    bce_weight: 5.0
+    dice_weight: 5.0
+  grasp_loss:
+    object_weight: 2.0
+    not_object_weight: 0.1
+    pseudo_ce_weight: 0.0
+    bce_topk: 512
+    bce_weight: 5.0
+    dice_weight: 5.0
+    deep_supervision: true
+    recompute_indices: true
+    adds_pred2gt: 100.0
+    adds_gt2pred: 0.0
+    adds_per_obj: false
+    contact_dir: 0.0
+    approach_dir: 0.0
+    offset: 1.0
+    offset_bin_weights:
+    - 0.16652107
+    - 0.21488856
+    - 0.37031708
+    - 0.55618503
+    - 0.75124664
+    - 0.93943357
+    - 1.07824539
+    - 1.19423112
+    - 1.55731375
+    - 3.17161779
+    gripper_name: robotiq_2f_140
+  place_loss:
+    bce_topk: 1024
+    bce_weight: 5.0
+    dice_weight: 5.0
+    deep_supervision: true
+diffusion:
+  obs_backbone: pointnet
+  checkpoint_object_encoder_pretrained: null
+  num_embed_dim: 256
+  num_obs_dim: 512
+  diffusion_embed_dim: 512
+  image_size: 256
+  num_diffusion_iters: 10
+  num_diffusion_iters_eval: 10
+  compositional_schedular: true
+  loss_pointmatching: false
+  loss_l1_pos: true
+  loss_l1_rot: true
+  grasp_repr: r3_so3
+  kappa: 2.02217
+  clip_sample: true
+  beta_schedule: squaredcos_cap_v2
+  npn_architecture: ffn
+  attention: cat_attn
+  gripper_name: robotiq_2f_140
+  pose_repr: mlp
+  num_grasps_per_object: 500
+  ptv3:
+    grid_size: 0.01
+discriminator:
+  checkpoint: graspgen_robotiq_2f_140_dis.pth 
+  checkpoint_object_encoder_pretrained: null
+  obs_backbone: pointnet
+  num_embed_dim: 256
+  num_obs_dim: 512
+  grasp_repr: r3_so3
+  pose_repr: mlp
+  topk_ratio: 0.75
+  kappa: 2.02217
+  gripper_name: robotiq_2f_140
+  ptv3:
+    grid_size: 0.01
+optimizer:
+  type: ADAMW
+  base_batch_size: 128
+  base_lr: 0.0008
+  backbone_multiplier: 1.0
+  grad_clip: -1
+  weight_decay: 0.05
+  lr_scale_multiplier: 1.0
+  lr: 5.0e-05
+  momentum: 0.9
+train:
+  model_name: m2t2
+  num_gpus: 1
+  port: '1234'
+  mask_thresh: 0.0
+  batch_size: 8
+  num_workers: 8
+  num_epochs: 160
+  print_freq: 25
+  plot_freq: 50
+  save_freq: 10
+  eval_freq: 5
+  checkpoint: null
+  log_dir: ''
+  num_scenes: null
+  debug: false
+  debug_ddp: false
+eval:
+  model_name: diffusion-discriminator
+  task: pick
+  split: valid
+  first_scene: -1
+  last_scene: 1000
+  scene: null
+  acronym_dir: ''
+  checkpoint: graspgen_robotiq_2f_140_gen.pth 
+  output_dir: null
+  exp_name: null
+  cam_coord: false
+  object_thresh: 0.0
+  mask_thresh: 0.6
+  seed_thresh: null
+  num_seed_grasps: 150
+  max_num_grasps: 200
+  retract: 0.0
+  grid_res: 0.01
+  placement_height: 0.02
+  placement_vis_radius: 0.2
+  batch_size: 1
+  num_workers: 1
+  print_freq: 1
+  num_procs: 4
+  num_vis: 5
+  num_runs: 1
+  debug: true
+  scene_file: ''
+  output_file: ''
+meshcat:
+  visualize: true
+  point_size: 0.01
+  line_width: 2
+obj:
+  scale: 1.0
+  num_sample_points: 2024

inference_manifest.json

@@ -3,15 +3,24 @@
     "digest": {
       "algorithm": "sha256",
       "scope": "structure",
-      "value": "b0a4edd6424e49ee3502079604ecc4fd972b59acd2f527d8af672da09c81b599"
+      "value": "3a71ce928384178e8a5b95bdb1214a6919553ba53fde9ca5bb14594dddad9bdc"
     },
     "files": [
+      {
+        "path": "assets/graspgen_config.yml"
+      },
+      {
+        "path": "assets/generator_checkpoint.pth"
+      },
+      {
+        "path": "assets/discriminator_checkpoint.pth"
+      },
       {
         "path": "assets/adapter.json"
       }
     ],
-    "name": "graspgen",
-    "revision": 4,
+    "name": "grasp",
+    "revision": 5,
     "uuid": "b300dc60-3bd4-4a64-8f70-670618235cbe"
   },
   "deployments": {
@@ -372,6 +381,24 @@
         }
       },
       "uuid": "11d63596-aa79-414d-ad3c-b63c4057fbd1"
+    },
+    "torch_cuda": {
+      "execution_contract": {
+        "cancellation_granularity": "request_boundary",
+        "execution_structure": "direct",
+        "state_scope": "request"
+      },
+      "revision": 1,
+      "runtime_profile": {
+        "backend": "torch",
+        "profile": {
+          "device": "cuda"
+        },
+        "target": {
+          "runtime": "torch"
+        }
+      },
+      "uuid": "2bbc4acc-5ce8-44c9-aa2d-74666d06f4f2"
     }
   },
   "model": {

inference_manifest.schema-v2.backup.json

@@ -0,0 +1,405 @@
+{
+  "bundle": {
+    "digest": {
+      "algorithm": "sha256",
+      "scope": "structure",
+      "value": "c2fa61f0f0cfc818104380861ca54a97de3c058696075965e6a9e31ce49506c0"
+    },
+    "files": [
+      {
+        "path": "assets/adapter.json"
+      },
+      {
+        "path": "assets/artifact-digests.json"
+      }
+    ],
+    "name": "grasp",
+    "revision": 3,
+    "uuid": "b300dc60-3bd4-4a64-8f70-670618235cbe"
+  },
+  "deployments": {
+    "ascend_310p": {
+      "artifacts": {
+        "denoiser": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/denoiser.om"
+        },
+        "discriminator_encoder_head": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/discriminator_encoder_head.om"
+        },
+        "discriminator_head": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/discriminator_head.om"
+        },
+        "discriminator_sa1": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/discriminator_sa1.om"
+        },
+        "discriminator_sa2": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/discriminator_sa2.om"
+        },
+        "generator_encoder_head": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/generator_encoder_head.om"
+        },
+        "generator_sa1": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/generator_sa1.om"
+        },
+        "generator_sa2": {
+          "format": "om",
+          "path": "artifacts/ascend/ascend_310p/generator_sa2.om"
+        }
+      },
+      "backend": "ascend",
+      "bindings": {
+        "denoiser": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "object_embedding",
+              "semantic": "internal.graspgen.generator_embedding",
+              "shape": [
+                1,
+                512
+              ]
+            },
+            {
+              "dtype": "float32",
+              "index": 1,
+              "runtime_name": "sample",
+              "semantic": "host.graspgen.diffusion_sample",
+              "shape": [
+                1000,
+                6
+              ]
+            },
+            {
+              "dtype": "float32",
+              "index": 2,
+              "runtime_name": "timestep",
+              "semantic": "host.graspgen.diffusion_time",
+              "shape": [
+                1
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/prediction_head/prediction_head.4/Gemm_add_130:0:predicted_noise",
+              "semantic": "host.graspgen.predicted_noise",
+              "shape": [
+                1000,
+                6
+              ]
+            }
+          ]
+        },
+        "discriminator_encoder_head": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "layout": "NCHW",
+              "runtime_name": "grouped_features",
+              "semantic": "host.graspgen.global_features",
+              "shape": [
+                1,
+                259,
+                1,
+                64
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/prediction_head/prediction_head.4/Gemm_add_17:0:object_embedding",
+              "semantic": "internal.graspgen.discriminator_embedding",
+              "shape": [
+                1,
+                512
+              ]
+            }
+          ]
+        },
+        "discriminator_head": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "object_embedding",
+              "semantic": "internal.graspgen.discriminator_embedding",
+              "shape": [
+                1,
+                512
+              ]
+            },
+            {
+              "dtype": "float32",
+              "index": 1,
+              "runtime_name": "grasp_rt",
+              "semantic": "host.graspgen.grasp_rt",
+              "shape": [
+                1000,
+                6
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/prediction_head/prediction_head.4/Gemm_add_24:0:logits",
+              "semantic": "host.graspgen.grasp_logits",
+              "shape": [
+                1000,
+                1
+              ]
+            },
+            {
+              "dtype": "float32",
+              "index": 1,
+              "runtime_name": "/Sigmoid:0:confidence",
+              "semantic": "host.graspgen.grasp_scores",
+              "shape": [
+                1000,
+                1
+              ]
+            }
+          ]
+        },
+        "discriminator_sa1": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "layout": "NCHW",
+              "runtime_name": "grouped_features",
+              "semantic": "host.graspgen.grouped_points",
+              "shape": [
+                1,
+                3,
+                256,
+                64
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/ReduceMax_ReduceMax_1:0:features",
+              "semantic": "host.graspgen.features",
+              "shape": [
+                1,
+                128,
+                256
+              ]
+            }
+          ]
+        },
+        "discriminator_sa2": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "layout": "NCHW",
+              "runtime_name": "grouped_features",
+              "semantic": "host.graspgen.stage2_features",
+              "shape": [
+                1,
+                131,
+                64,
+                128
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/ReduceMax_ReduceMax_1:0:features",
+              "semantic": "host.graspgen.features",
+              "shape": [
+                1,
+                256,
+                64
+              ]
+            }
+          ]
+        },
+        "generator_encoder_head": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "layout": "NCHW",
+              "runtime_name": "grouped_features",
+              "semantic": "host.graspgen.global_features",
+              "shape": [
+                1,
+                259,
+                1,
+                64
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/prediction_head/prediction_head.4/Gemm_add_17:0:object_embedding",
+              "semantic": "internal.graspgen.generator_embedding",
+              "shape": [
+                1,
+                512
+              ]
+            }
+          ]
+        },
+        "generator_sa1": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "layout": "NCHW",
+              "runtime_name": "grouped_features",
+              "semantic": "host.graspgen.grouped_points",
+              "shape": [
+                1,
+                3,
+                256,
+                64
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/ReduceMax_ReduceMax_1:0:features",
+              "semantic": "host.graspgen.features",
+              "shape": [
+                1,
+                128,
+                256
+              ]
+            }
+          ]
+        },
+        "generator_sa2": {
+          "inputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "layout": "NCHW",
+              "runtime_name": "grouped_features",
+              "semantic": "host.graspgen.stage2_features",
+              "shape": [
+                1,
+                131,
+                64,
+                128
+              ]
+            }
+          ],
+          "outputs": [
+            {
+              "dtype": "float32",
+              "index": 0,
+              "runtime_name": "PartitionedCall_/ReduceMax_ReduceMax_1:0:features",
+              "semantic": "host.graspgen.features",
+              "shape": [
+                1,
+                256,
+                64
+              ]
+            }
+          ]
+        }
+      },
+      "device_links": [
+        {
+          "consumer": "denoiser",
+          "lifetime": "inference",
+          "owner": "producer",
+          "producer": "generator_encoder_head",
+          "producer_binding": "output",
+          "semantic": "internal.graspgen.generator_embedding",
+          "transport": "device_pointer"
+        },
+        {
+          "consumer": "discriminator_head",
+          "lifetime": "inference",
+          "owner": "producer",
+          "producer": "discriminator_encoder_head",
+          "producer_binding": "output",
+          "semantic": "internal.graspgen.discriminator_embedding",
+          "transport": "device_pointer"
+        }
+      ],
+      "execution": [
+        "generator_sa1",
+        "generator_sa2",
+        "generator_encoder_head",
+        "discriminator_sa1",
+        "discriminator_sa2",
+        "discriminator_encoder_head",
+        "denoiser",
+        "discriminator_head"
+      ],
+      "revision": 1,
+      "target": {
+        "runtime": "acl",
+        "soc": "Ascend310P1"
+      },
+      "uuid": "11d63596-aa79-414d-ad3c-b63c4057fbd1"
+    }
+  },
+  "model": {
+    "family": "graspgen",
+    "inputs": [
+      {
+        "dtype": "float32",
+        "semantic": "observation.object_points",
+        "shape": [
+          -1,
+          3
+        ]
+      }
+    ],
+    "kind": "perception",
+    "operation": "",
+    "outputs": [
+      {
+        "dtype": "float32",
+        "semantic": "grasp.poses",
+        "shape": [
+          -1,
+          4,
+          4
+        ]
+      },
+      {
+        "dtype": "float32",
+        "semantic": "grasp.confidence",
+        "shape": [
+          -1
+        ]
+      }
+    ],
+    "semantic_identity": {
+      "logical_model_revision": "graspgen@v2",
+      "output_semantics": "grasp-pose-matrix-confidence-sorted-v1",
+      "preprocessing_contract": "object-points-outlier-filtered-centered-kappa-scaled-v1"
+    }
+  },
+  "schema_version": 2
+}