Organize checkpoints by four experimental settings

README.md

@@ -10,10 +10,21 @@ tags:
   - importance-sampling
 ---
 
-# ReTime / OPTR — Florence–Bridge and Qwen–Bridge handoff
+# ReTime / OPTR — four-setting handoff
 
 This package contains the server-authoritative Florence–Bridge OPTR code, its complete 20K common-start state and Uniform/Residual-OPTR 20.5K evaluation snapshots. It also contains the complete Qwen3-VL-4B–Bridge 29K common-start state and the environment/seed contract needed for paired multi-seed continuation experiments.
 
+## Unified checkpoint layout
+
+The release is being normalized under [`settings/`](settings/README.md). Each of the four settings has the same entries: `common_start`, `uniform`, `residual`, `keyframe`, `frameskip`, `loss_is`, `configs`, and `dependencies`.
+
+- [`settings/florence_bridge/`](settings/florence_bridge/README.md)
+- [`settings/florence_robotwin/`](settings/florence_robotwin/README.md)
+- [`settings/qwen_bridge/`](settings/qwen_bridge/README.md)
+- [`settings/qwen_robotwin/`](settings/qwen_robotwin/README.md)
+
+The common-start entry is the resumable branch point. Method endpoints are evaluation-ready model snapshots. Shared source code is kept once at repository root and referenced by each setting to avoid four inconsistent copies.
+
 ## Rebuttal 训练与测评种子
 
 | 设置 | Common-start 训练 seed | Continuation 训练 seed | Evaluation policy seeds |

settings/ASSET_STATUS.md

@@ -0,0 +1,10 @@
+# Asset status
+
+The organized tree is populated in two stages:
+
+- Existing Hub objects are copied server-side into their setting directories.
+- Missing common-start or Uniform/Residual artifacts are uploaded directly from the server that owns the authoritative checkpoint.
+
+Expected final entries for every setting are `common_start`, `uniform`, `residual`, `keyframe`, `frameskip`, `loss_is`, `configs`, and `dependencies`.
+
+`common_start` means a resumable training state. The five method directories contain evaluation-ready endpoint weights. Training configs, proposal/cache metadata and environment notes live beside the corresponding setting and point to shared code rather than duplicating the whole source tree four times.

settings/README.md

@@ -0,0 +1,21 @@
+# Four-setting checkpoint index
+
+This directory groups the release assets by experimental setting. Each setting uses the same six logical entries:
+
+1. `common_start/`: the branch point used by all continuation methods;
+2. `uniform/`: the Uniform continuation checkpoint;
+3. `residual/`: the Residual-OPTR (ReTime) continuation checkpoint;
+4. `keyframe/`: the KF-Weight continuation checkpoint;
+5. `frameskip/`: the FrameSkip continuation checkpoint;
+6. `loss_is/`: the active Loss-IS continuation checkpoint.
+
+The `common_start` entry is the resumable training state. Method endpoints are evaluation-ready model snapshots unless a setting README explicitly lists additional training-state files. Shared source code remains at repository root; each setting README records the corresponding step, model family, environment and method dependencies.
+
+| Setting | Common start | Continuation endpoint |
+|---|---:|---:|
+| Florence–Bridge | 20,000 | 20,500 |
+| Florence–RoboTwin | 10,000 | 10,500 |
+| Qwen3-VL-4B–Bridge | 29,000 | 29,500 |
+| Qwen3-VL-4B–RoboTwin | 11,000 | 11,500 |
+
+No implementation-path or self-referential checksum gate is required for launch. Reproduction should still validate the model family, dataset, episode/anchor support, proposal semantics, batch contract and training budget.

settings/florence_bridge/README.md

@@ -0,0 +1,17 @@
+# Florence–Bridge
+
+- Backbone: Florence-2-large with DiT-B flow-matching action head.
+- Common start: step 20,000; full two-rank resumable state.
+- Continuation endpoint: step 20,500 (+500 optimizer updates).
+- Main continuation seed: `117465`; common-start seed: `42`.
+- Evaluation: SimplerEnv WidowX, tasks Stack/Carrot/Spoon/Eggplant, policy seeds `518957, 765353, 686765, 516408`, 24 episodes per task and seed, no video.
+- Training contract: two GPUs, BF16, ZeRO-2, micro-batch 4/GPU, gradient accumulation 8, global batch 64.
+
+Method endpoints under this directory are model-only evaluation snapshots. Resume new branches from `common_start/steps_20000_accelerate/`.
+
+Method dependencies:
+
+- Residual: frozen K=8, alpha=0.2 temporal proposal and ordinary p/q correction.
+- FrameSkip: FrameSkip cache and adapter under the shared external-baseline source tree.
+- Keyframe: KF-Weight/APE cache and adapter under the shared external-baseline source tree.
+- Loss-IS: active always-on candidate sampling implementation; candidate factor 3 and ordinary importance correction.

settings/florence_bridge/common_start/steps_20000_accelerate/checkpoint.json

@@ -0,0 +1,48 @@
+{
+  "completed_steps": 20000,
+  "files": [
+    {
+      "bytes": 13,
+      "name": "latest"
+    },
+    {
+      "bytes": 5514449457,
+      "name": "pytorch_model/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt"
+    },
+    {
+      "bytes": 5514463601,
+      "name": "pytorch_model/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt"
+    },
+    {
+      "bytes": 1839089981,
+      "name": "pytorch_model/mp_rank_00_model_states.pt"
+    },
+    {
+      "bytes": 2154184469,
+      "name": "pytorch_model.pt"
+    },
+    {
+      "bytes": 15017,
+      "name": "random_states_0.pkl"
+    },
+    {
+      "bytes": 15081,
+      "name": "random_states_1.pkl"
+    },
+    {
+      "bytes": 33270,
+      "name": "zero_to_fp32.py"
+    }
+  ],
+  "resume_identity": {
+    "config_sha256": "1c394b6043c323133b0512d3efd761ae8332e5098969864f410b71651aa56f44",
+    "manifest_sha256": "a1627df8ff110c7298eaf025ffcd4f48d9ca17117cfcfb3926b962d0bb19c0ca",
+    "model_structure_sha256": "e055e3fa0085edc6a9bf6cdb1605a05e61ae289957e24e37cfe2cfce2dc94c39",
+    "proposal_alpha": null,
+    "proposal_sha256": null,
+    "seed": 42,
+    "total_batch_size": 64
+  },
+  "schema_version": 1,
+  "world_size": 2
+}

settings/florence_bridge/common_start/steps_20000_accelerate/latest

@@ -0,0 +1 @@
+pytorch_model

settings/florence_bridge/common_start/steps_20000_accelerate/pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4f39cd339271347b1ab83ddc9c8dae7e07e315bea8cf86812b2e0b6772abd42
+size 2154184469

settings/florence_bridge/common_start/steps_20000_accelerate/pytorch_model/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42263278dd16848dd50fd159f430bb40efa541d8b62aba5d608531d3af90f1d4
+size 5514449457

settings/florence_bridge/common_start/steps_20000_accelerate/pytorch_model/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6e21653b454605b3cb3db62bf9b1bae90ddf63eea1389cc0a0dd351f1f0839c
+size 5514463601

settings/florence_bridge/common_start/steps_20000_accelerate/pytorch_model/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:822f82ffa3ebb7f8e1e9cb15f7c24dafb4054d9fb2b63646fb81ba63172c09b6
+size 1839089981

settings/florence_bridge/common_start/steps_20000_accelerate/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29e1b8cf5838519c544dfe660c71c064b59cd35d8f5c5ae9fa1e63e9cfd3e773
+size 15017

settings/florence_bridge/common_start/steps_20000_accelerate/random_states_1.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fed527bb36c43639d19411c7da8f3eb3eca04543ed7eb06f62c1fa978f3bdf31
+size 15081

settings/florence_bridge/common_start/steps_20000_accelerate/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

settings/florence_bridge/configs/deepspeed_florence_2gpu.yaml

@@ -0,0 +1,15 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_config_file: configs/ds_florence_zero2.json
+distributed_type: DEEPSPEED
+downcast_bf16: "no"
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+use_cpu: false

settings/florence_bridge/configs/ds_florence_zero2.json

@@ -0,0 +1,19 @@
+{
+  "fp16": {"enabled": false},
+  "bf16": {"enabled": true},
+  "train_micro_batch_size_per_gpu": "auto",
+  "train_batch_size": "auto",
+  "gradient_accumulation_steps": "auto",
+  "zero_optimization": {
+    "stage": 2,
+    "allgather_partitions": true,
+    "allgather_bucket_size": 50000000,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 50000000,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "cpu_offload": false
+  },
+  "gradient_clipping": 1.0,
+  "steps_per_print": 10
+}

settings/florence_bridge/configs/florence_bridge_20k_plus500.yaml

@@ -0,0 +1,110 @@
+run_id: florence-bridge-20k-plus500
+run_root_dir: ${oc.env:OPTR_RUN_ROOT}
+seed: 117465
+wandb_entity: null
+wandb_project: OPTR-Florence-Bridge
+is_debug: false
+version_id: "0.21"
+
+framework:
+  name: QwenGR00T
+  qwenvl:
+    base_vlm: ${oc.env:FLORENCE_MODEL_PATH}
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+  dino:
+    dino_backbone: dinov2_vits14
+  action_model:
+    action_model_type: DiT-B
+    action_hidden_dim: 1024
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 16
+    future_action_window_size: 15
+    past_action_window_size: 0
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    data_root_dir: ${oc.env:BRIDGE_DATA_ROOT}
+    data_mix: bridge_rt_infoflow
+    action_type: delta_ee
+    action_mode: abs
+    precomputed_statistics_path: ${oc.env:BRIDGE_STATISTICS_PATH}
+    precomputed_statistics_key_map:
+      bridge_orig_lerobot: oxe_bridge
+      fractal20220817_data_lerobot: oxe_rt1
+    CoT_prompt: Your task is {instruction}. To identify the key objects for your task. Locate their bounding boxes in [x1,y1,x2,y2] format.
+    per_device_batch_size: 4
+    load_all_data_for_training: true
+    obs_image_size: [224, 224]
+    video_backend: torchvision_av
+    num_workers: 4
+    balance_dataset_weights: false
+    balance_trajectory_weights: false
+    traceable_sampling: true
+    expected_world_size: 2
+    dataset_provenance_path: ${oc.env:BRIDGE_PROVENANCE_PATH}
+    dataset_content_fingerprint_path: ${oc.env:BRIDGE_FINGERPRINT_PATH}
+    valid_anchor_manifest_path: ${oc.env:BRIDGE_MANIFEST_PATH}
+    temporal_proposal_path: null
+    transform_rng_seed: 298226
+
+trainer:
+  use_deepspeed: true
+  max_train_steps: 100000
+  num_warmup_steps: 5000
+  save_interval: null
+  save_steps: [20500]
+  evaluation_snapshot_interval: null
+  eval_interval: 100001
+  repeated_diffusion_steps: 4
+  pretrained_checkpoint: null
+  is_resume: true
+  allow_model_only_resume: false
+  learning_rate:
+    base: 3.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 1.0e-04
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+  freeze_modules: ""
+  trainable_modules: ""
+  frozen_modules_eval: false
+  logging_frequency: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  save_format: pt
+  save_final_checkpoint: false
+  optimizer:
+    name: AdamW
+    betas: [0.9, 0.95]
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  resume_accelerate_checkpoint: ${oc.env:COMMON_START_20K}
+  branch_parent_seed: 42
+  branch_rng_seed: 117465
+  model_rng_seed: 588643
+  stop_after_steps: 20500
+
+output_dir: ${oc.env:OPTR_RUN_ROOT}/florence-bridge-20k-plus500

settings/florence_bridge/configs/reference/common_start_20k_full.yaml

@@ -0,0 +1,113 @@
+run_id: florence-scratch-uniform-seed-42-official100k-mb4-eval1k-20260726
+run_root_dir: /mnt/miaochengyu/starvla-infoflow-runtime/runs/florence-scratch-official-100k
+seed: 42
+wandb_entity: null
+wandb_project: InfoFlow-StarVLA-BridgeRT
+is_debug: false
+version_id: '0.21'
+framework:
+  name: QwenGR00T
+  qwenvl:
+    base_vlm: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-2-large
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+  action_model:
+    action_model_type: DiT-B
+    action_hidden_dim: 1024
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 16
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+    future_action_window_size: 15
+    past_action_window_size: 0
+  dino:
+    dino_backbone: dinov2_vits14
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    data_root_dir: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Datasets/OXE_LEROBOT_DATASET
+    data_mix: bridge_rt_infoflow
+    action_type: delta_ee
+    action_mode: abs
+    precomputed_statistics_path: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-GR00T-Bridge-RT-1/dataset_statistics.json
+    precomputed_statistics_key_map:
+      bridge_orig_lerobot: oxe_bridge
+      fractal20220817_data_lerobot: oxe_rt1
+    CoT_prompt: Your task is {instruction}. To identify the key objects for your task.
+      Locate their bounding boxes in [x1,y1,x2,y2] format.
+    per_device_batch_size: 4
+    load_all_data_for_training: true
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    num_workers: 4
+    balance_dataset_weights: false
+    balance_trajectory_weights: false
+    traceable_sampling: true
+    expected_world_size: 2
+    dataset_provenance_path: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/data_provenance.json
+    dataset_content_fingerprint_path: /mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/content_fingerprints.json
+    valid_anchor_manifest_path: /mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/valid_anchor_manifest_florence_scratch_v1.json
+    temporal_proposal_path: null
+trainer:
+  use_deepspeed: true
+  max_train_steps: 100000
+  num_warmup_steps: 5000
+  save_interval: null
+  save_steps:
+  - 5000
+  - 10000
+  - 20000
+  - 30000
+  - 50000
+  - 75000
+  - 100000
+  evaluation_snapshot_interval: 1000
+  eval_interval: 100001
+  repeated_diffusion_steps: 4
+  pretrained_checkpoint: null
+  is_resume: false
+  allow_model_only_resume: false
+  learning_rate:
+    base: 3.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+  freeze_modules: ''
+  trainable_modules: ''
+  frozen_modules_eval: false
+  logging_frequency: 50
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  save_format: pt
+  save_final_checkpoint: false
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+config_yaml: examples/InfoFlow/train_files/server/bridgert_florence_scratch_official_100k_mb4_eval1k.yaml
+output_dir: /mnt/miaochengyu/starvla-infoflow-runtime/runs/florence-scratch-official-100k/florence-scratch-uniform-seed-42-official100k-mb4-eval1k-20260726

settings/florence_bridge/configs/reference/residual_seed117465_full.yaml

@@ -0,0 +1,118 @@
+run_id: infoflow-confirmatory-residual-seed-117465
+run_root_dir: /mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-confirmatory-22k-20260730
+seed: 117465
+wandb_entity: null
+wandb_project: InfoFlow-StarVLA-BridgeRT
+is_debug: false
+version_id: '0.21'
+framework:
+  name: QwenGR00T
+  qwenvl:
+    base_vlm: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-2-large
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+  action_model:
+    action_model_type: DiT-B
+    action_hidden_dim: 1024
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 16
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+    future_action_window_size: 15
+    past_action_window_size: 0
+  dino:
+    dino_backbone: dinov2_vits14
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    data_root_dir: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Datasets/OXE_LEROBOT_DATASET
+    data_mix: bridge_rt_infoflow
+    action_type: delta_ee
+    action_mode: abs
+    precomputed_statistics_path: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-GR00T-Bridge-RT-1/dataset_statistics.json
+    precomputed_statistics_key_map:
+      bridge_orig_lerobot: oxe_bridge
+      fractal20220817_data_lerobot: oxe_rt1
+    CoT_prompt: Your task is {instruction}. To identify the key objects for your task.
+      Locate their bounding boxes in [x1,y1,x2,y2] format.
+    per_device_batch_size: 4
+    load_all_data_for_training: true
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    num_workers: 4
+    balance_dataset_weights: false
+    balance_trajectory_weights: false
+    traceable_sampling: true
+    expected_world_size: 2
+    dataset_provenance_path: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/data_provenance.json
+    dataset_content_fingerprint_path: /mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/content_fingerprints.json
+    valid_anchor_manifest_path: /mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/valid_anchor_manifest_florence_scratch_v1.json
+    valid_anchor_manifest_migration_path: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/valid_anchor_manifest_rng_migration_20260729.json
+    temporal_proposal_path: /mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-proposals/scratch20k_residual_grid8_gate42f_batch1_20260729/residual_temporal_proposal.json
+    transform_rng_seed: 298226
+trainer:
+  use_deepspeed: true
+  max_train_steps: 100000
+  num_warmup_steps: 5000
+  save_interval: null
+  save_steps:
+  - 20500
+  - 21000
+  - 22000
+  evaluation_snapshot_interval: 1000
+  eval_interval: 100001
+  repeated_diffusion_steps: 4
+  pretrained_checkpoint: null
+  is_resume: true
+  allow_model_only_resume: false
+  learning_rate:
+    base: 3.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+  freeze_modules: ''
+  trainable_modules: ''
+  frozen_modules_eval: false
+  logging_frequency: 50
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  save_format: pt
+  save_final_checkpoint: false
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  resume_accelerate_checkpoint: /mnt/miaochengyu/starvla-infoflow-runtime/runs/florence-scratch-official-100k/florence-scratch-uniform-seed-42-official100k-mb4-eval1k-20260726/checkpoints/steps_20000_accelerate
+  resume_checkpoint_integrity_manifest: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/common_start_accelerate_integrity_20260730.json
+  resume_checkpoint_integrity_manifest_sha256: 899d2cef3f73191e11f7a1628afc05b60c80af5816b8fd77c1c69e90d5c9ff39
+  branch_parent_seed: 42
+  branch_rng_seed: 117465
+  model_rng_seed: 588643
+  stop_after_steps: 22000
+config_yaml: /home/miaochengyu/starVLA-infoflow-codex/examples/InfoFlow/train_files/server/generated/infoflow_confirmatory_20k_20260730/confirmatory_residual_seed_117465.yaml
+output_dir: /mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-confirmatory-22k-20260730/infoflow-confirmatory-residual-seed-117465

settings/florence_bridge/configs/reference/uniform_seed117465_full.yaml

@@ -0,0 +1,118 @@
+run_id: infoflow-confirmatory-uniform-seed-117465
+run_root_dir: /mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-confirmatory-22k-20260730
+seed: 117465
+wandb_entity: null
+wandb_project: InfoFlow-StarVLA-BridgeRT
+is_debug: false
+version_id: '0.21'
+framework:
+  name: QwenGR00T
+  qwenvl:
+    base_vlm: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-2-large
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+  action_model:
+    action_model_type: DiT-B
+    action_hidden_dim: 1024
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 16
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+    future_action_window_size: 15
+    past_action_window_size: 0
+  dino:
+    dino_backbone: dinov2_vits14
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    data_root_dir: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Datasets/OXE_LEROBOT_DATASET
+    data_mix: bridge_rt_infoflow
+    action_type: delta_ee
+    action_mode: abs
+    precomputed_statistics_path: /mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-GR00T-Bridge-RT-1/dataset_statistics.json
+    precomputed_statistics_key_map:
+      bridge_orig_lerobot: oxe_bridge
+      fractal20220817_data_lerobot: oxe_rt1
+    CoT_prompt: Your task is {instruction}. To identify the key objects for your task.
+      Locate their bounding boxes in [x1,y1,x2,y2] format.
+    per_device_batch_size: 4
+    load_all_data_for_training: true
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    num_workers: 4
+    balance_dataset_weights: false
+    balance_trajectory_weights: false
+    traceable_sampling: true
+    expected_world_size: 2
+    dataset_provenance_path: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/data_provenance.json
+    dataset_content_fingerprint_path: /mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/content_fingerprints.json
+    valid_anchor_manifest_path: /mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/valid_anchor_manifest_florence_scratch_v1.json
+    valid_anchor_manifest_migration_path: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/valid_anchor_manifest_rng_migration_20260729.json
+    temporal_proposal_path: null
+    transform_rng_seed: 298226
+trainer:
+  use_deepspeed: true
+  max_train_steps: 100000
+  num_warmup_steps: 5000
+  save_interval: null
+  save_steps:
+  - 20500
+  - 21000
+  - 22000
+  evaluation_snapshot_interval: 1000
+  eval_interval: 100001
+  repeated_diffusion_steps: 4
+  pretrained_checkpoint: null
+  is_resume: true
+  allow_model_only_resume: false
+  learning_rate:
+    base: 3.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+  freeze_modules: ''
+  trainable_modules: ''
+  frozen_modules_eval: false
+  logging_frequency: 50
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  save_format: pt
+  save_final_checkpoint: false
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  resume_accelerate_checkpoint: /mnt/miaochengyu/starvla-infoflow-runtime/runs/florence-scratch-official-100k/florence-scratch-uniform-seed-42-official100k-mb4-eval1k-20260726/checkpoints/steps_20000_accelerate
+  resume_checkpoint_integrity_manifest: /home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/common_start_accelerate_integrity_20260730.json
+  resume_checkpoint_integrity_manifest_sha256: 899d2cef3f73191e11f7a1628afc05b60c80af5816b8fd77c1c69e90d5c9ff39
+  branch_parent_seed: 42
+  branch_rng_seed: 117465
+  model_rng_seed: 588643
+  stop_after_steps: 22000
+config_yaml: /home/miaochengyu/starVLA-infoflow-codex/examples/InfoFlow/train_files/server/generated/infoflow_confirmatory_20k_20260730/confirmatory_uniform_seed_117465.yaml
+output_dir: /mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-confirmatory-22k-20260730/infoflow-confirmatory-uniform-seed-117465

settings/florence_bridge/frameskip/steps_20500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48667f000750716c82b5ee1cebf67f8b6ab1c282d5b06588485daf577d7c0701
+size 2154184469

settings/florence_bridge/keyframe/steps_20500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:560ba3cc4ff1649280d7acb559f7392f4b611c4416849c5cc17e236092c6f6cb
+size 2154184469

settings/florence_bridge/loss_is/steps_20500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c12d8e09042f3b24720c0bf7345e8facb920d8db095804c7a24ff25f34a8d36
+size 2154184469

settings/florence_bridge/residual/steps_20500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d443c8d9cc3caf1ffab4d75a9e35d849d79ed1abe01626e8a93757d64cb3d97e
+size 2154184469

settings/florence_bridge/uniform/steps_20500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e363c4a70f64868f1cad77479f3f06f906fd3fe47a2f40174fd78afafff32147
+size 2154184469

settings/florence_robotwin/README.md

@@ -0,0 +1,17 @@
+# Florence–RoboTwin
+
+- Backbone: Florence-2-large with the RoboTwin three-view, 14D action interface.
+- Common start: step 10,000; full two-rank resumable state.
+- Continuation endpoint: step 10,500 (+500 optimizer updates).
+- Continuation seed: `484370`.
+- Evaluation: the fixed 20-task RoboTwin panel, Easy (`demo_clean`) and Hard (`demo_randomized`), 25 episodes per task and condition.
+- Training contract: two GPUs and the server-authoritative Florence–RoboTwin batch/optimizer configuration shipped with the setting assets.
+
+Method endpoints are model-only evaluation snapshots. Resume new branches from `common_start/steps_10000_accelerate/`.
+
+Method dependencies:
+
+- Residual: RoboTwin K=8 residual proposal, 14D trajectory support and ordinary p/q correction.
+- FrameSkip: RoboTwin cache; the selected physical timestep is applied consistently across all camera modalities.
+- Keyframe: RoboTwin 14D KF cache and action-statistics adapter.
+- Loss-IS: active Loss-IS candidate sampler with the RoboTwin model interface and paired flow-noise/time reuse.

settings/florence_robotwin/frameskip/steps_10500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03745fecff8f7e98da348cb95d38ed27e840f4d6856d63b299a3468e30e73c55
+size 2152621885

settings/florence_robotwin/keyframe/steps_10500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f2fcde1201f78dd169c9da1b2c07cc8e5e54720a5d613f9e34dffb2dd194010
+size 2152621885

settings/florence_robotwin/loss_is/steps_10500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2b0d1c85f74ac1d72b2f6704e3d6acd1fd7595a5801629d6f96ba2246028ec7
+size 2152621885

settings/qwen_bridge/README.md

@@ -0,0 +1,16 @@
+# Qwen3-VL-4B–Bridge
+
+- Backbone: Qwen3-VL-4B-Instruct with DiT-B flow-matching action head.
+- Common start: step 29,000; paired model and training-state files are both required for exact continuation.
+- Continuation endpoint: step 29,500 (+500 optimizer updates).
+- Common-start training seed: `42`; the historical continuation inherits the saved 29K RNG/data-order state.
+- Evaluation: SimplerEnv WidowX, tasks Stack/Carrot/Spoon/Eggplant, policy seeds `518957, 765353, 686765, 516408`, 24 episodes per task and seed, no video.
+
+Method endpoints are model-only evaluation snapshots. Exact continuation from 29K requires both files under `common_start/` plus the included configuration and environment description.
+
+Method dependencies:
+
+- Residual: Qwen–Bridge K=8 residual proposal and ordinary p/q correction.
+- FrameSkip: Qwen–Bridge FrameSkip cache and adapter.
+- Keyframe: Qwen–Bridge KF cache and 7D action-statistics adapter.
+- Loss-IS: active candidate sampling, candidate factor 3, replacement sampling and ordinary importance correction.

settings/qwen_bridge/common_start/steps_29000_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d23a05fbe874a7cf5e100378c29aa043dd4594e2777a10c6306d32c493f35f61
+size 9199045796

settings/qwen_bridge/common_start/steps_29000_training_state.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd81e7a88a5a4fc06ed8d445e002d00a236e16385d617c94cf101e6f944cf276
+size 9346430417

settings/qwen_bridge/dependencies/AGENTS.md

@@ -0,0 +1,26 @@
+# Qwen–Bridge 29K Common Start 操作约束
+
+本目录用于从固定 29K common start 开展 continuation 实验。操作前阅读 `README.md`、`ENVIRONMENT.md`、`MULTISEED_PROTOCOL.md` 和 [Bridge 训练与测评种子总表](https://huggingface.co/mmmmmcyaaa/ReTime/blob/main/BRIDGE_SEED_PROTOCOL.md)。
+
+## 不可改变
+
+- 起点必须同时加载 `steps_29000_pytorch_model.pt` 与 `steps_29000_training_state.pt`。
+- common-start 历史训练 seed 是 `42`。
+- 每个主动 reseed realization 下，各比较方法必须共享同一个 continuation training seed。
+- 新的 Qwen paired continuation seed 清单待定。
+- 29K 与 29.5K checkpoint 的 Rebuttal 测评固定使用 policy seeds `518957, 765353, 686765, 516408`。
+- 每个 checkpoint 的正式测评是 4 seeds × 4 tasks × 24 episodes = 384 rollouts。
+- global batch、学习率、scheduler、训练预算、world size、数据与 support 合同必须在方法间保持一致。
+
+## 禁止
+
+- 禁止把 evaluation seed 当成 training seed。
+- 区分历史 `seed=42` 的完整恢复与未来主动 reseed。
+- 禁止只恢复模型而丢弃 optimizer、scheduler、RNG 或 data-order。
+- 禁止新增代码、实现、绝对路径、manifest 自指等工程哈希启动门。
+
+## 允许的可移植性修改
+
+- 可修改本机数据、基础模型与输出目录的绝对路径。
+- 可按目标硬件调整启动器，但不得改变方法间的 global batch 或训练预算。
+- 可记录普通传输审计信息，但不得让它成为训练启动条件。

settings/qwen_bridge/dependencies/ENVIRONMENT.md

@@ -0,0 +1,72 @@
+# Qwen–Bridge 29K Common Start 环境
+
+## 系统与硬件
+
+- Ubuntu 22.04
+- Linux kernel 6.8.0-138
+- NVIDIA driver 570.211.01
+- 历史训练硬件：RTX 4090
+- Python 3.11.13
+
+## 核心软件
+
+| 包 | 版本 |
+|---|---|
+| PyTorch | 2.7.1+cu126 |
+| CUDA runtime | 12.6 |
+| cuDNN | 9.5.0.1 |
+| torchvision | 0.22.1 |
+| transformers | 4.57.0 |
+| accelerate | 1.13.0 |
+| bitsandbytes | 0.50.0 |
+| deepspeed | 0.19.2（本历史 run 未启用） |
+| flash-attn | 2.7.4.post1 |
+| huggingface-hub | 0.36.2 |
+| omegaconf | 2.3.0 |
+| peft | 0.19.1 |
+| safetensors | 0.5.3 |
+| tokenizers | 0.22.2 |
+| wandb | 0.23.1 |
+| numpy | 1.26.4 |
+| pandas | 2.2.3 |
+| polars | 1.30.0 |
+| datasets | 3.6.0 |
+| lerobot | 0.1.0 |
+| einops | 0.8.1 |
+| timm | 1.0.24 |
+| av | 17.0.0 |
+
+## 外部依赖（不重复上传）
+
+- Base VLM：`Qwen3-VL-4B-Instruct`。需单独从官方模型源取得 tokenizer、processor 和基础权重。
+- Bridge/OXE 数据：`bridge_rt_infoflow` 对应的 OXE LeRobot 数据及 statistics。
+- StarVLA/OPTR 源码：使用交接仓库中与本配置匹配的 QwenGR00T、DiT-B action head、dataset 和 trainer 实现。
+
+## 历史训练合同
+
+- framework：QwenGR00T。
+- action model：DiT-B，16 layers，dropout 0.2。
+- action/state dim：7/7。
+- action horizon：16。
+- BF16。
+- 单卡训练。
+- per-device batch：2。
+- gradient accumulation：24。
+- global batch：48。
+- optimizer：AdamW8bit，betas=(0.9, 0.95)，eps=1e-8，weight decay=1e-8。
+- base LR：3e-5；VLM interface LR：1e-5；action model LR：1e-4。
+- cosine scheduler，minimum LR=5e-7，100K scheduler horizon，10K warmup。
+- gradient clipping：1.0。
+- 日志频率：每 10 updates。
+- 29K checkpoint 训练时全部约 4.599B 参数可训练，`freeze_modules` 为空。
+
+## 恢复要求
+
+必须同时加载：
+
+1. `steps_29000_pytorch_model.pt`；
+2. `steps_29000_training_state.pt`。
+
+后者包含 optimizer、scheduler、RNG 与 data-order counter。只载入模型权重不是完整 common-start 恢复。
+
+不得新增 implementation/code/path/current-manifest SHA 启动门。启动前只核对文件存在且可加载、step=29000、数据与 support 语义一致、global batch 与训练预算正确。

settings/qwen_bridge/dependencies/MULTISEED_PROTOCOL.md

@@ -0,0 +1,9 @@
+# Qwen–Bridge continuation 与测评种子
+
+Florence 和 Qwen 的 Rebuttal 训练、测评种子统一登记在 [Bridge 训练与测评种子总表](https://huggingface.co/mmmmmcyaaa/ReTime/blob/main/BRIDGE_SEED_PROTOCOL.md)。
+
+29K common start 的基础训练 seed 为 `42`。历史 29K → 29.5K continuation 配置为 `seed: 42`，继承 checkpoint 中保存的 RNG 与数据抽取位置。
+
+Rebuttal 若新增主动 reseed 的 Qwen continuation，训练 seed 清单待定。同一组的配对运行共享 continuation seed，并以相同方式初始化 Python、NumPy、PyTorch CPU/CUDA、dataloader、采样和模型 RNG。模型、optimizer、scheduler、起始 step、batch、学习率和训练预算沿用共同协议。
+
+29K 与 29.5K checkpoint 的 Rebuttal 测评统一使用 `518957, 765353, 686765, 516408`。每组评测为四个 policy seeds × Stack、Carrot、Spoon、Eggplant 四任务 × 24 episodes，共 384 rollouts，BF16、无视频。

settings/qwen_bridge/dependencies/config.full.yaml

@@ -0,0 +1,122 @@
+run_id: qwen3vl-groot-bridge-commonstart-scratch-15k-seed42
+run_root_dir: /data/yanfei/qwen_bridge_commonstart_20k_to30k_20260911/runs
+seed: 42
+wandb_entity: null
+wandb_project: QwenBridge-CommonStart
+is_debug: false
+version_id: '0.21'
+framework:
+  name: QwenGR00T
+  qwenvl:
+    base_vlm: /home/yanfei/miaochengyu/robotwin-infoflow/legacy_bridge_runtime/compat/mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: DiT-B
+    action_hidden_dim: 2048
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 16
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 2560
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+    future_action_window_size: 15
+    past_action_window_size: 0
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    data_root_dir: /data/yanfei/qwen_bridge_assets/OXE_LEROBOT_DATASET
+    data_mix: bridge_rt_infoflow
+    action_type: delta_ee
+    action_mode: abs
+    precomputed_statistics_path: /home/yanfei/miaochengyu/robotwin-infoflow/legacy_bridge_runtime/compat/mnt/miaochengyu/starvla-infoflow-runtime/playground/Pretrained_models/Florence-GR00T-Bridge-RT-1/dataset_statistics.json
+    precomputed_statistics_key_map:
+      bridge_orig_lerobot: oxe_bridge
+      fractal20220817_data_lerobot: oxe_rt1
+    CoT_prompt: Your task is {instruction}. To identify the key objects for your task.
+      Locate their bounding boxes in [x1,y1,x2,y2] format.
+    per_device_batch_size: 2
+    load_all_data_for_training: true
+    obs:
+    - image_0
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    num_workers: 1
+    balance_dataset_weights: false
+    balance_trajectory_weights: false
+    traceable_sampling: true
+    expected_world_size: 1
+    dataset_provenance_path: /home/yanfei/miaochengyu/robotwin-infoflow/legacy_bridge_runtime/compat/home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/data_provenance.json
+    dataset_content_fingerprint_path: /home/yanfei/miaochengyu/robotwin-infoflow/legacy_bridge_runtime/compat/mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/content_fingerprints.json
+    valid_anchor_manifest_path: /home/yanfei/miaochengyu/robotwin-infoflow/legacy_bridge_runtime/compat/mnt/miaochengyu/starvla-infoflow-runtime/manifests/bridgert/valid_anchor_manifest_florence_scratch_v1.json
+    valid_anchor_manifest_migration_path: /home/yanfei/miaochengyu/robotwin-infoflow/legacy_bridge_runtime/compat/home/miaochengyu/starVLA-infoflow-codex/protocols/bridgert/valid_anchor_manifest_rng_migration_20260729.json
+    temporal_proposal_path: null
+trainer:
+  use_deepspeed: false
+  model_dtype: bfloat16
+  max_train_steps: 100000
+  stop_after_steps: 15000
+  initial_completed_steps: 0
+  num_warmup_steps: 10000
+  repeated_diffusion_steps: 4
+  evaluation_snapshot_interval: null
+  save_interval: null
+  save_steps:
+  - 21000
+  - 22000
+  - 23000
+  - 24000
+  - 25000
+  - 26000
+  - 27000
+  - 28000
+  - 29000
+  - 30000
+  save_final_checkpoint: false
+  eval_interval: null
+  pretrained_checkpoint: null
+  is_resume: true
+  allow_model_only_resume: false
+  learning_rate:
+    base: 3.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+  freeze_modules: ''
+  trainable_modules: ''
+  frozen_modules_eval: false
+  logging_frequency: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 24
+  save_format: pt
+  optimizer:
+    name: AdamW8bit
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  operational_stop_after_steps: 30000
+config_yaml: /home/yanfei/miaochengyu/qwen_bridge_commonstart_4090_20260910/configs/qwen_bridge_commonstart_scratch_15k.yaml
+output_dir: /data/yanfei/qwen_bridge_commonstart_20k_to30k_20260911/runs/qwen3vl-groot-bridge-commonstart-scratch-15k-seed42

settings/qwen_bridge/dependencies/requirements-qwen-bridge.txt

@@ -0,0 +1,21 @@
+torch==2.7.1
+torchvision==0.22.1
+transformers==4.57.0
+accelerate==1.13.0
+bitsandbytes==0.50.0
+deepspeed==0.19.2
+flash-attn==2.7.4.post1
+huggingface-hub==0.36.2
+omegaconf==2.3.0
+peft==0.19.1
+safetensors==0.5.3
+tokenizers==0.22.2
+wandb==0.23.1
+numpy==1.26.4
+pandas==2.2.3
+polars==1.30.0
+datasets==3.6.0
+lerobot==0.1.0
+einops==0.8.1
+timm==1.0.24
+av==17.0.0

settings/qwen_bridge/frameskip/steps_29500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bb2fdbc798a64ee8d648ae55ed7bd620279e1165db39da73b2d4d34c0ab24cb
+size 9199045796

settings/qwen_bridge/keyframe/steps_29500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d68f460d5eae83e38be8ca10958f624d25b99939212ba933ce4d0c8fe6470604
+size 9199045796

settings/qwen_bridge/loss_is/steps_29500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26ec40656e644374fac557c19e9344681391989d448cfc023e9391f02f1b7068
+size 9199045796

settings/qwen_robotwin/README.md

@@ -0,0 +1,17 @@
+# Qwen3-VL-4B–RoboTwin
+
+- Backbone: Qwen3-VL-4B-Instruct with the RoboTwin three-view, 14D action interface.
+- Common start: step 11,000; full two-rank ZeRO resumable state.
+- Continuation endpoint: step 11,500 (+500 optimizer updates).
+- Continuation seed: `484370`; transform seed `289726`; model seed `721927`.
+- Evaluation: the fixed 20-task RoboTwin panel, Easy (`demo_clean`) and Hard (`demo_randomized`), 25 episodes per task and condition.
+- Training contract: two GPUs, per-device batch 5, gradient accumulation 4, global batch 40.
+
+Method endpoints are consolidated model-only evaluation snapshots. Resume new branches from `common_start/steps_11000_accelerate/`.
+
+Method dependencies:
+
+- Residual: RoboTwin K=8 residual proposal, 14D support and ordinary p/q correction.
+- FrameSkip: RoboTwin cache with synchronized temporal remapping for all camera views.
+- Keyframe: 14D KF cache/action-statistics implementation.
+- Loss-IS: active candidate sampling with paired flow-noise/time reuse and the Qwen–RoboTwin model interface.

settings/qwen_robotwin/frameskip/steps_11500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:018797073168650b9db654c04be89818f8bd188821ffdb29122a427497755ad9
+size 18394366619

settings/qwen_robotwin/keyframe/steps_11500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7adca3f3032ab4e105e0b9cda7b54cf6052fbcb5e038d56bc413dc42c5cb0fcb
+size 18394384980

settings/qwen_robotwin/loss_is/steps_11500_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9219baf845811fe6a7f4ddccf762e289aed7fa3a6873bf2d25b8d3e3091d8cbc
+size 18394384980