Upload folder using huggingface_hub

.gitattributes

@@ -42,3 +42,5 @@ flappy_fix_latency_3_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024
 flappy_fix_latency_3_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123423-tte0csme/run-tte0csme.wandb filter=lfs diff=lfs merge=lfs -text
 deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024002-tdvnpwth/run-tdvnpwth.wandb filter=lfs diff=lfs merge=lfs -text
 deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123335-0b3r9ly8/run-0b3r9ly8.wandb filter=lfs diff=lfs merge=lfs -text
+demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/run-61qjyp27.wandb filter=lfs diff=lfs merge=lfs -text
+demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/run-3aeecfjc.wandb filter=lfs diff=lfs merge=lfs -text

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ce7b40220c1f0406652f40e00394d708cbae33ddf08d59cb8bca8c129ebc753
+size 9784896838

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state/latest

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

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state/pytorch_model/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c80aa20bea39a3cbd2f3075673a4735c4300e80432a57a4edc106eaca1502f18
+size 9007152504

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b390ddf3fbe806b45e0fc898f8e191f0159a1ec86215c9dfa474c0be64dae79b
+size 14604

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state/random_states_1.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28f01d59656fdd858f84676b30750d7b3c90fbeaa8d581a5bf35f1991d9b2727
+size 14604

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state/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 not ZERO_STAGE 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(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"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)

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.full.yaml

@@ -0,0 +1,257 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    flex_backend: triton
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 6
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+    data_mix: demon_attack_train__bridge
+    eval_data_mix: demon_attack_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 8
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: demon_attack_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: demon_attack_train
+  mixed_converted_name: demon_attack_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: raw_frames
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 4000
+  num_warmup_steps: 100
+  save_interval: 500
+  eval_interval: 250
+  eval_num_batches: 50
+  per_latency_eval_num_batches: null
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: deepspeed
+  is_resume: true
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
+  resume_step: 2000
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    image_transform: raw_rgb
+    prompt_mode: raw
+    ghost_trail:
+      history_frames: 5
+      gamma: 1.3
+      min_alpha: 35
+      scroll_px_per_step: 4.0
+      ground_fraction: 0.22
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: 'You are playing Demon Attack from a single game image. Choose
+      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+    eval_parallel_envs: 5
+    action_chunk_execution:
+      enabled: false
+      chunk_size: null
+    demon_attack:
+      noop_max: 30
+    enabled: true
+    eval_backend: latency_bench
+    distributed_mode: rank_sharded
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 250
+      latencies:
+      - 2
+      num_episodes: 20
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 2
+      num_episodes: 50
+      max_steps_per_episode: 3600
+  task: demon_attack
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: demon_attack
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_training_state: true
+  save_safetensors_file: true
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+config_yaml: null
+is_debug: false
+version_id: '0.21'

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.yaml

@@ -0,0 +1,89 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_safetensors_file: true
+  save_training_state: true
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    latency_curriculum:
+      enabled: false
+    per_device_batch_size: 8
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 6
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+  kv_memory:
+    enabled: false
+    packed_train: false
+    rebased_sink: true
+    rollout_len: 8
+    window: 4
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+rl_games:
+  env_eval:
+    enabled: true
+    eval_backend: latency_bench
+    mid_train:
+      enabled: false
+      interval_steps: 250
+    prompt_mode: raw
+  task: demon_attack
+run_id: demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: deepspeed
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  eval_interval: 250
+  eval_num_batches: 50
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 16
+  is_resume: true
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 4000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: true
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: null
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
+  profile_timing:
+    enabled: false
+  resume_step: 2000
+  save_interval: 500
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.05213047191500664,
+        0.25790393352508545,
+        0.17633378505706787,
+        0.08325807005167007,
+        0.18504540622234344,
+        0.24532835185527802,
+        0.0
+      ],
+      "std": [
+        0.22280143201351166,
+        0.4365144968032837,
+        0.3812168538570404,
+        0.2774599492549896,
+        0.3870716392993927,
+        0.4308818578720093,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 1055141,
+    "num_trajectories": 198
+  }
+}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.06258254498243332,
+        0.28070709109306335,
+        0.17078126966953278,
+        0.07629787921905518,
+        0.1742354929447174,
+        0.23539571464061737,
+        0.0
+      ],
+      "std": [
+        0.2421988993883133,
+        0.44936150312423706,
+        0.37632739543914795,
+        0.26546531915664673,
+        0.379315584897995,
+        0.42424726486206055,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 9843,
+    "num_trajectories": 2
+  }
+}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/config.yaml

@@ -0,0 +1,255 @@
+framework:
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    flex_backend: triton
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 6
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+    data_mix: demon_attack_train__bridge
+    eval_data_mix: demon_attack_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 8
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: demon_attack_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: demon_attack_train
+  mixed_converted_name: demon_attack_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: raw_frames
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 4000
+  num_warmup_steps: 100
+  save_interval: 500
+  eval_interval: 250
+  eval_num_batches: 50
+  per_latency_eval_num_batches: null
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: deepspeed
+  is_resume: true
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
+  resume_step: 2000
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: ${oc.env:WANDB_ENTITY}
+wandb_project: ${oc.env:WANDB_PROJECT,starVLA_rl_games}
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    image_transform: raw_rgb
+    prompt_mode: raw
+    ghost_trail:
+      history_frames: 5
+      gamma: 1.3
+      min_alpha: 35
+      scroll_px_per_step: 4.0
+      ground_fraction: 0.22
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: 'You are playing Demon Attack from a single game image. Choose
+      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+    eval_parallel_envs: 5
+    action_chunk_execution:
+      enabled: false
+      chunk_size: null
+    demon_attack:
+      noop_max: 30
+    enabled: true
+    eval_backend: latency_bench
+    distributed_mode: rank_sharded
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 250
+      latencies:
+      - 2
+      num_episodes: 20
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 2
+      num_episodes: 50
+      max_steps_per_episode: 3600
+  task: demon_attack
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: demon_attack
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_training_state: true
+  save_safetensors_file: true
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: 0.21

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml

@@ -0,0 +1,375 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=demon_attack
+    - init=bridge
+    - mode=single
+    - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++framework.qwenvl.attn_implementation=flash_attention_2
+    - ++framework.qwenvl.flex_backend=triton
+    - ++framework.qwenvl.enable_gradient_checkpointing=true
+    - ++framework.action_model.state_dim=7
+    - ++framework.action_model.loss_type=discrete_ce
+    - ++framework.action_model.action_horizon=1
+    - ++framework.action_model.future_action_window_size=0
+    - ++framework.action_model.past_action_window_size=0
+    - ++framework.action_model.action_dim=7
+    - ++framework.action_model.action_env_dim=6
+    - ++framework.kv_memory.enabled=false
+    - ++framework.kv_memory.window=4
+    - ++framework.kv_memory.rollout_len=8
+    - ++framework.kv_memory.packed_train=false
+    - ++framework.kv_memory.rebased_sink=true
+    - ++framework.name=QwenOFT
+    - ++datasets.vla_data.dataset_py=lerobot_datasets
+    - ++datasets.vla_data.include_state=true
+    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+    - ++datasets.vla_data.data_mix=demon_attack_train
+    - ++datasets.vla_data.eval_data_mix=null
+    - ++datasets.vla_data.custom_mixtures_path=null
+    - ++datasets.vla_data.action_type=discrete
+    - ++datasets.vla_data.sequential_step_sampling=false
+    - ++datasets.vla_data.eval_sequential_step_sampling=null
+    - ++datasets.vla_data.num_workers=8
+    - ++datasets.vla_data.eval_num_workers=8
+    - ++datasets.vla_data.prefetch_factor=4
+    - ++datasets.vla_data.persistent_workers=true
+    - ++datasets.vla_data.pin_memory=true
+    - ++datasets.vla_data.shuffle=true
+    - ++datasets.vla_data.action_balance.enabled=false
+    - ++datasets.vla_data.action_balance.strategy=balanced_epoch
+    - ++datasets.vla_data.action_balance.action_key=action_id
+    - ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+    - ++datasets.vla_data.action_balance.noop_id=0
+    - ++datasets.vla_data.action_balance.flap_id=1
+    - ++datasets.vla_data.latency_curriculum.enabled=false
+    - ++datasets.vla_data.latency_curriculum.strategy=exclusive
+    - ++datasets.vla_data.latency_curriculum.latencies=null
+    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+    - ++datasets.vla_data.per_device_batch_size=8
+    - ++datasets.vla_data.load_all_data_for_training=true
+    - ++datasets.vla_data.num_obs_frames=1
+    - ++datasets.vla_data.image_mode=single
+    - ++datasets.vla_data.prompt_mode=raw
+    - ++datasets.vla_data.stitch_grid=[2,2]
+    - ++datasets.vla_data.obs_image_size=null
+    - ++datasets.vla_data.video_backend=torchvision_av
+    - ++dataset.source_hf=
+    - ++dataset.config_name=null
+    - ++dataset.source_subdir=null
+    - ++dataset.converted_name=demon_attack_train
+    - ++dataset.single_source_hf=
+    - ++dataset.mixed_source_hf=
+    - ++dataset.single_converted_name=demon_attack_train
+    - ++dataset.mixed_converted_name=demon_attack_mixed_latency_train
+    - ++dataset.single_latency_filter=null
+    - ++dataset.mixed_latency_filter=null
+    - ++dataset.force_download=false
+    - ++dataset.setup_force=false
+    - ++dataset.skip_verification=false
+    - ++dataset.target_latency_unit=raw_frames
+    - ++dataset.verify_rows=200
+    - ++dataset.max_episodes=null
+    - ++dataset.episodes_per_latency=null
+    - ++dataset.latency_filter=null
+    - ++dataset.debug_subset.enabled=false
+    - ++dataset.debug_subset.max_episodes=5
+    - ++dataset.debug_subset.suffix=debug
+    - ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+    - ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+    - ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+    - ++trainer.max_train_steps=4000
+    - ++trainer.num_warmup_steps=100
+    - ++trainer.save_interval=500
+    - ++trainer.eval_interval=250
+    - ++trainer.eval_num_batches=50
+    - ++trainer.per_latency_eval_num_batches=null
+    - ++trainer.eval_action_classification=false
+    - ++trainer.eval_action_classification_interval=null
+    - ++trainer.cc_f1_tolerance=1
+    - ++trainer.learning_rate.base=2e-05
+    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+    - ++trainer.learning_rate.action_model=0.0001
+    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - ++trainer.freeze_modules=
+    - ++trainer.freeze_vit=false
+    - ++trainer.freeze_tied_embedding=false
+    - ++trainer.freeze_llm_layers=[]
+    - ++trainer.loss_scale.vla=1.0
+    - ++trainer.loss_scale.vlm=0.1
+    - ++trainer.max_grad_norm=1.0
+    - ++trainer.weight_decay=0.0
+    - ++trainer.logging_frequency=1
+    - ++trainer.profile_timing.enabled=false
+    - ++trainer.profile_timing.log_interval=10
+    - ++trainer.gradient_clipping=1.0
+    - ++trainer.gradient_accumulation_steps=16
+    - ++trainer.distributed_backend=deepspeed
+    - ++trainer.is_resume=false
+    - ++trainer.pretrained_checkpoint=null
+    - ++trainer.resume_step=0
+    - ++trainer.reload_modules=null
+    - ++trainer.optimizer.name=AdamW
+    - ++trainer.optimizer.betas=[0.9,0.95]
+    - ++trainer.optimizer.eps=1e-08
+    - ++trainer.optimizer.weight_decay=1e-08
+    - ++trainer.optimizer.fused=true
+    - ++trainer.save_format=pt
+    - ++workspace_dir=WORKSPACE_DIR
+    - ++run_root_dir=results/Checkpoints
+    - ++seed=42
+    - ++auth.env_file=null
+    - ++auth.hf_token_env=HF_TOKEN
+    - ++auth.wandb_api_key_env=WANDB_API_KEY
+    - ++paths.run_root_dir=results/Checkpoints
+    - ++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+    - ++paths.dataset_cache_dir=null
+    - ++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+    - ++rl_games.model_alias=openvla
+    - ++rl_games.env_eval.image_size=224
+    - ++rl_games.env_eval.frameskip=4
+    - ++rl_games.env_eval.image_transform=raw_rgb
+    - ++rl_games.env_eval.prompt_mode=raw
+    - ++rl_games.env_eval.ghost_trail.history_frames=5
+    - ++rl_games.env_eval.ghost_trail.gamma=1.3
+    - ++rl_games.env_eval.ghost_trail.min_alpha=35
+    - ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+    - ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+    - ++rl_games.env_eval.seed=42
+    - ++rl_games.env_eval.fixed_episode_seeds=true
+    - ++rl_games.env_eval.latency_seed_stride=0
+    - ++rl_games.env_eval.task_seed_stride=0
+    - '++rl_games.env_eval.task_description=''You are playing Demon Attack from a
+      single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT,
+      RIGHTFIRE, LEFTFIRE.'''
+    - ++rl_games.env_eval.eval_parallel_envs=5
+    - ++rl_games.env_eval.action_chunk_execution.enabled=false
+    - ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+    - ++rl_games.env_eval.demon_attack.noop_max=30
+    - ++rl_games.env_eval.enabled=true
+    - ++rl_games.env_eval.eval_backend=latency_bench
+    - ++rl_games.env_eval.distributed_mode=rank_sharded
+    - ++rl_games.env_eval.vectorized.enabled=false
+    - ++rl_games.env_eval.vectorized.batch_size=1
+    - ++rl_games.env_eval.latency.prompt_map_path=null
+    - ++rl_games.env_eval.latency.mode=single
+    - ++rl_games.env_eval.latency.values=[0]
+    - ++rl_games.env_eval.mid_train.enabled=false
+    - ++rl_games.env_eval.mid_train.interval_steps=250
+    - ++rl_games.env_eval.mid_train.latencies=[2]
+    - ++rl_games.env_eval.mid_train.num_episodes=20
+    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - ++rl_games.env_eval.post_train.enabled=false
+    - ++rl_games.env_eval.post_train.latencies=[2]
+    - ++rl_games.env_eval.post_train.num_episodes=50
+    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - ++rl_games.task=demon_attack
+    - ++rl_games.initialization_mode=bridge
+    - ++rl_games.action_carrier=bridge
+    - ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    - ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    - ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++checkpoint.load=auto
+    - ++checkpoint.hf_repo_id=null
+    - ++checkpoint.save_best_model=false
+    - ++checkpoint.save_final_model=true
+    - ++checkpoint.save_pt_file=false
+    - ++checkpoint.save_training_state=true
+    - ++checkpoint.save_safetensors_file=true
+    - ++checkpoint.local.keep_last_n=1
+    - ++checkpoint.sync.enabled=false
+    - ++checkpoint.sync.repo_id=null
+    - ++checkpoint.sync.keep_last_n=0
+    - ++checkpoint.sync.sync_every_n_checkpoints=1
+    - ++checkpoint.sync.resume_policy=local_latest
+    - ++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+    - ++output_dir=null
+    - ++config_yaml=null
+    - ++is_debug=false
+    - ++version_id=0.21
+    - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+    - ++trainer.is_resume=true
+    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
+    - ++trainer.resume_step=2000
+    - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+    - ++datasets.vla_data.data_mix=demon_attack_train__bridge
+    - ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+    - ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - ++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: '++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action,++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action,++checkpoint.hf_repo_id=null,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=false,++checkpoint.save_safetensors_file=true,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=false,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=null,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=null,++dataset.converted_name=demon_attack_train,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=null,++dataset.mixed_converted_name=demon_attack_mixed_latency_train,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=demon_attack_train,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=,++dataset.source_subdir=null,++dataset.target_latency_unit=raw_frames,++dataset.verify_rows=200,++datasets.vla_data.action_balance.action_key=action_id,++datasets.vla_data.action_balance.enabled=false,++datasets.vla_data.action_balance.flap_id=1,++datasets.vla_data.action_balance.noop_id=0,++datasets.vla_data.action_balance.strategy=balanced_epoch,++datasets.vla_data.action_balance.target_flap_fraction=0.3,++datasets.vla_data.action_type=discrete,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=demon_attack_train,++datasets.vla_data.data_mix=demon_attack_train__bridge,++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=single,++datasets.vla_data.include_state=true,++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=1,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=null,++datasets.vla_data.per_device_batch_size=8,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=false,++datasets.vla_data.shuffle=true,++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=6,++framework.action_model.action_horizon=1,++framework.action_model.future_action_window_size=0,++framework.action_model.loss_type=discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=QwenOFT,++framework.qwenvl.attn_implementation=flash_attention_2,++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.enable_gradient_checkpointing=true,++framework.qwenvl.flex_backend=triton,++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1,++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps,++paths.run_root_dir=results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.env_eval.action_chunk_execution.chunk_size=null,++rl_games.env_eval.action_chunk_execution.enabled=false,++rl_games.env_eval.demon_attack.noop_max=30,++rl_games.env_eval.distributed_mode=rank_sharded,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=latency_bench,++rl_games.env_eval.eval_parallel_envs=5,++rl_games.env_eval.fixed_episode_seeds=true,++rl_games.env_eval.frameskip=4,++rl_games.env_eval.ghost_trail.gamma=1.3,++rl_games.env_eval.ghost_trail.ground_fraction=0.22,++rl_games.env_eval.ghost_trail.history_frames=5,++rl_games.env_eval.ghost_trail.min_alpha=35,++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0,++rl_games.env_eval.image_size=224,++rl_games.env_eval.image_transform=raw_rgb,++rl_games.env_eval.latency.mode=single,++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json,++rl_games.env_eval.latency.prompt_map_path=null,++rl_games.env_eval.latency.values=[0],++rl_games.env_eval.latency_seed_stride=0,++rl_games.env_eval.mid_train.enabled=false,++rl_games.env_eval.mid_train.interval_steps=250,++rl_games.env_eval.mid_train.latencies=[2],++rl_games.env_eval.mid_train.max_steps_per_episode=3600,++rl_games.env_eval.mid_train.num_episodes=20,++rl_games.env_eval.post_train.enabled=false,++rl_games.env_eval.post_train.latencies=[2],++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=50,++rl_games.env_eval.prompt_mode=raw,++rl_games.env_eval.seed=42,++rl_games.env_eval.task_description=''You
+      are playing Demon Attack from a single game image. Choose exactly one action
+      from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'',++rl_games.env_eval.task_seed_stride=0,++rl_games.env_eval.vectorized.batch_size=1,++rl_games.env_eval.vectorized.enabled=false,++rl_games.initialization_mode=bridge,++rl_games.model_alias=openvla,++rl_games.task=demon_attack,++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline,++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints,++run_root_dir=results/Checkpoints,++seed=42,++trainer.cc_f1_tolerance=1,++trainer.distributed_backend=deepspeed,++trainer.eval_action_classification=false,++trainer.eval_action_classification_interval=null,++trainer.eval_interval=250,++trainer.eval_num_batches=50,++trainer.freeze_llm_layers=[],++trainer.freeze_modules=,++trainer.freeze_tied_embedding=false,++trainer.freeze_vit=false,++trainer.gradient_accumulation_steps=16,++trainer.gradient_clipping=1.0,++trainer.is_resume=false,++trainer.is_resume=true,++trainer.learning_rate.action_model=0.0001,++trainer.learning_rate.base=2e-05,++trainer.learning_rate.qwen_vl_interface=1e-05,++trainer.logging_frequency=1,++trainer.loss_scale.vla=1.0,++trainer.loss_scale.vlm=0.1,++trainer.lr_scheduler_type=cosine_with_min_lr,++trainer.max_grad_norm=1.0,++trainer.max_train_steps=4000,++trainer.num_warmup_steps=100,++trainer.optimizer.betas=[0.9,0.95],++trainer.optimizer.eps=1e-08,++trainer.optimizer.fused=true,++trainer.optimizer.name=AdamW,++trainer.optimizer.weight_decay=1e-08,++trainer.per_latency_eval_num_batches=null,++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state,++trainer.pretrained_checkpoint=null,++trainer.profile_timing.enabled=false,++trainer.profile_timing.log_interval=10,++trainer.reload_modules=null,++trainer.resume_step=0,++trainer.resume_step=2000,++trainer.save_format=pt,++trainer.save_interval=500,++trainer.scheduler_specific_kwargs.min_lr=1e-06,++trainer.weight_decay=0.0,++version_id=0.21,++workspace_dir=WORKSPACE_DIR,env=demon_attack,init=bridge,mode=single,model=openvla'
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.4
+    version_base: '1.1'
+    cwd: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: demon_attack
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml

@@ -0,0 +1,213 @@
+- model=openvla
+- env=demon_attack
+- init=bridge
+- mode=single
+- ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++framework.qwenvl.attn_implementation=flash_attention_2
+- ++framework.qwenvl.flex_backend=triton
+- ++framework.qwenvl.enable_gradient_checkpointing=true
+- ++framework.action_model.state_dim=7
+- ++framework.action_model.loss_type=discrete_ce
+- ++framework.action_model.action_horizon=1
+- ++framework.action_model.future_action_window_size=0
+- ++framework.action_model.past_action_window_size=0
+- ++framework.action_model.action_dim=7
+- ++framework.action_model.action_env_dim=6
+- ++framework.kv_memory.enabled=false
+- ++framework.kv_memory.window=4
+- ++framework.kv_memory.rollout_len=8
+- ++framework.kv_memory.packed_train=false
+- ++framework.kv_memory.rebased_sink=true
+- ++framework.name=QwenOFT
+- ++datasets.vla_data.dataset_py=lerobot_datasets
+- ++datasets.vla_data.include_state=true
+- ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+- ++datasets.vla_data.data_mix=demon_attack_train
+- ++datasets.vla_data.eval_data_mix=null
+- ++datasets.vla_data.custom_mixtures_path=null
+- ++datasets.vla_data.action_type=discrete
+- ++datasets.vla_data.sequential_step_sampling=false
+- ++datasets.vla_data.eval_sequential_step_sampling=null
+- ++datasets.vla_data.num_workers=8
+- ++datasets.vla_data.eval_num_workers=8
+- ++datasets.vla_data.prefetch_factor=4
+- ++datasets.vla_data.persistent_workers=true
+- ++datasets.vla_data.pin_memory=true
+- ++datasets.vla_data.shuffle=true
+- ++datasets.vla_data.action_balance.enabled=false
+- ++datasets.vla_data.action_balance.strategy=balanced_epoch
+- ++datasets.vla_data.action_balance.action_key=action_id
+- ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+- ++datasets.vla_data.action_balance.noop_id=0
+- ++datasets.vla_data.action_balance.flap_id=1
+- ++datasets.vla_data.latency_curriculum.enabled=false
+- ++datasets.vla_data.latency_curriculum.strategy=exclusive
+- ++datasets.vla_data.latency_curriculum.latencies=null
+- ++datasets.vla_data.latency_curriculum.phase_steps=null
+- ++datasets.vla_data.latency_curriculum.phase_distributions=null
+- ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+- ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+- ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+- ++datasets.vla_data.latency_curriculum.final_equalization=true
+- ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+- ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+- ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+- ++datasets.vla_data.latency_curriculum.computed_plan=null
+- ++datasets.vla_data.per_device_batch_size=8
+- ++datasets.vla_data.load_all_data_for_training=true
+- ++datasets.vla_data.num_obs_frames=1
+- ++datasets.vla_data.image_mode=single
+- ++datasets.vla_data.prompt_mode=raw
+- ++datasets.vla_data.stitch_grid=[2,2]
+- ++datasets.vla_data.obs_image_size=null
+- ++datasets.vla_data.video_backend=torchvision_av
+- ++dataset.source_hf=
+- ++dataset.config_name=null
+- ++dataset.source_subdir=null
+- ++dataset.converted_name=demon_attack_train
+- ++dataset.single_source_hf=
+- ++dataset.mixed_source_hf=
+- ++dataset.single_converted_name=demon_attack_train
+- ++dataset.mixed_converted_name=demon_attack_mixed_latency_train
+- ++dataset.single_latency_filter=null
+- ++dataset.mixed_latency_filter=null
+- ++dataset.force_download=false
+- ++dataset.setup_force=false
+- ++dataset.skip_verification=false
+- ++dataset.target_latency_unit=raw_frames
+- ++dataset.verify_rows=200
+- ++dataset.max_episodes=null
+- ++dataset.episodes_per_latency=null
+- ++dataset.latency_filter=null
+- ++dataset.debug_subset.enabled=false
+- ++dataset.debug_subset.max_episodes=5
+- ++dataset.debug_subset.suffix=debug
+- ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+- ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+- ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+- ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+- ++trainer.max_train_steps=4000
+- ++trainer.num_warmup_steps=100
+- ++trainer.save_interval=500
+- ++trainer.eval_interval=250
+- ++trainer.eval_num_batches=50
+- ++trainer.per_latency_eval_num_batches=null
+- ++trainer.eval_action_classification=false
+- ++trainer.eval_action_classification_interval=null
+- ++trainer.cc_f1_tolerance=1
+- ++trainer.learning_rate.base=2e-05
+- ++trainer.learning_rate.qwen_vl_interface=1e-05
+- ++trainer.learning_rate.action_model=0.0001
+- ++trainer.lr_scheduler_type=cosine_with_min_lr
+- ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+- ++trainer.freeze_modules=
+- ++trainer.freeze_vit=false
+- ++trainer.freeze_tied_embedding=false
+- ++trainer.freeze_llm_layers=[]
+- ++trainer.loss_scale.vla=1.0
+- ++trainer.loss_scale.vlm=0.1
+- ++trainer.max_grad_norm=1.0
+- ++trainer.weight_decay=0.0
+- ++trainer.logging_frequency=1
+- ++trainer.profile_timing.enabled=false
+- ++trainer.profile_timing.log_interval=10
+- ++trainer.gradient_clipping=1.0
+- ++trainer.gradient_accumulation_steps=16
+- ++trainer.distributed_backend=deepspeed
+- ++trainer.is_resume=false
+- ++trainer.pretrained_checkpoint=null
+- ++trainer.resume_step=0
+- ++trainer.reload_modules=null
+- ++trainer.optimizer.name=AdamW
+- ++trainer.optimizer.betas=[0.9,0.95]
+- ++trainer.optimizer.eps=1e-08
+- ++trainer.optimizer.weight_decay=1e-08
+- ++trainer.optimizer.fused=true
+- ++trainer.save_format=pt
+- ++workspace_dir=WORKSPACE_DIR
+- ++run_root_dir=results/Checkpoints
+- ++seed=42
+- ++auth.env_file=null
+- ++auth.hf_token_env=HF_TOKEN
+- ++auth.wandb_api_key_env=WANDB_API_KEY
+- ++paths.run_root_dir=results/Checkpoints
+- ++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+- ++paths.dataset_cache_dir=null
+- ++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+- ++rl_games.model_alias=openvla
+- ++rl_games.env_eval.image_size=224
+- ++rl_games.env_eval.frameskip=4
+- ++rl_games.env_eval.image_transform=raw_rgb
+- ++rl_games.env_eval.prompt_mode=raw
+- ++rl_games.env_eval.ghost_trail.history_frames=5
+- ++rl_games.env_eval.ghost_trail.gamma=1.3
+- ++rl_games.env_eval.ghost_trail.min_alpha=35
+- ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+- ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+- ++rl_games.env_eval.seed=42
+- ++rl_games.env_eval.fixed_episode_seeds=true
+- ++rl_games.env_eval.latency_seed_stride=0
+- ++rl_games.env_eval.task_seed_stride=0
+- '++rl_games.env_eval.task_description=''You are playing Demon Attack from a single
+  game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE,
+  LEFTFIRE.'''
+- ++rl_games.env_eval.eval_parallel_envs=5
+- ++rl_games.env_eval.action_chunk_execution.enabled=false
+- ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+- ++rl_games.env_eval.demon_attack.noop_max=30
+- ++rl_games.env_eval.enabled=true
+- ++rl_games.env_eval.eval_backend=latency_bench
+- ++rl_games.env_eval.distributed_mode=rank_sharded
+- ++rl_games.env_eval.vectorized.enabled=false
+- ++rl_games.env_eval.vectorized.batch_size=1
+- ++rl_games.env_eval.latency.prompt_map_path=null
+- ++rl_games.env_eval.latency.mode=single
+- ++rl_games.env_eval.latency.values=[0]
+- ++rl_games.env_eval.mid_train.enabled=false
+- ++rl_games.env_eval.mid_train.interval_steps=250
+- ++rl_games.env_eval.mid_train.latencies=[2]
+- ++rl_games.env_eval.mid_train.num_episodes=20
+- ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- ++rl_games.env_eval.post_train.enabled=false
+- ++rl_games.env_eval.post_train.latencies=[2]
+- ++rl_games.env_eval.post_train.num_episodes=50
+- ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+- ++rl_games.task=demon_attack
+- ++rl_games.initialization_mode=bridge
+- ++rl_games.action_carrier=bridge
+- ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+- ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+- ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++checkpoint.load=auto
+- ++checkpoint.hf_repo_id=null
+- ++checkpoint.save_best_model=false
+- ++checkpoint.save_final_model=true
+- ++checkpoint.save_pt_file=false
+- ++checkpoint.save_training_state=true
+- ++checkpoint.save_safetensors_file=true
+- ++checkpoint.local.keep_last_n=1
+- ++checkpoint.sync.enabled=false
+- ++checkpoint.sync.repo_id=null
+- ++checkpoint.sync.keep_last_n=0
+- ++checkpoint.sync.sync_every_n_checkpoints=1
+- ++checkpoint.sync.resume_policy=local_latest
+- ++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+- ++output_dir=null
+- ++config_yaml=null
+- ++is_debug=false
+- ++version_id=0.21
+- ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+- ++trainer.is_resume=true
+- ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
+- ++trainer.resume_step=2000
+- ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+- ++datasets.vla_data.data_mix=demon_attack_train__bridge
+- ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+- ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- ++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/summary.jsonl

@@ -0,0 +1,8 @@
+{"steps": 500}
+{"steps": 1000}
+{"steps": 1500}
+{"steps": 2000}
+{"steps": 2500}
+{"steps": 3000}
+{"steps": 3500}
+{"steps": 4000}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/debug.log

@@ -0,0 +1,25 @@
+2026-07-26 12:36:18,858 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Configure stats pid to 1083569
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug.log
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug-internal.log
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():768] calling init triggers
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():816] starting backend
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():831] sending inform_init request
+2026-07-26 12:36:19,204 INFO    MainThread:1083569 [wandb_init.py:init():836] backend started and connected
+2026-07-26 12:36:19,218 INFO    MainThread:1083569 [wandb_init.py:init():906] updated telemetry
+2026-07-26 12:36:19,250 INFO    MainThread:1083569 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-07-26 12:36:19,872 INFO    MainThread:1083569 [wandb_init.py:init():974] starting run threads in backend
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_console_start():2523] atexit reg
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-07-26 12:36:20,166 INFO    MainThread:1083569 [wandb_init.py:init():1012] run started, returning control to user process
+2026-07-26 16:16:06,499 INFO    MainThread:1083569 [wandb_run.py:_finish():2285] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/3aeecfjc
+2026-07-26 16:16:06,500 INFO    MainThread:1083569 [wandb_run.py:_atexit_cleanup():2490] got exitcode: 0
+2026-07-26 16:16:06,501 INFO    MainThread:1083569 [wandb_run.py:_restore():2472] restore
+2026-07-26 16:16:06,501 INFO    MainThread:1083569 [wandb_run.py:_restore():2478] restore done
+2026-07-26 16:16:08,181 INFO    MainThread:1083569 [wandb_run.py:_footer_sync_info():3895] logging synced files

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260725_220307-jb36xg06/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb":{"runtime":0},"_runtime":0}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260725_220307-jb36xg06/logs/debug-core.log

@@ -0,0 +1,25 @@
+{"time":"2026-07-25T22:01:29.60209943-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmprowg5pf0/port-352860.txt","pid":352860,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-07-25T22:01:29.60330571-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":352860}
+{"time":"2026-07-25T22:01:29.603060248-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-352860-353050-535436166/socket","Net":"unix"}}
+{"time":"2026-07-25T22:01:29.753050176-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-07-25T22:02:23.867311282-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"2(@)"}
+{"time":"2026-07-25T22:02:24.007439775-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"3(@)"}
+{"time":"2026-07-25T22:03:07.875599169-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"jb36xg06","id":"2(@)"}
+{"time":"2026-07-25T22:03:08.360170889-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"jb36xg06","id":"2(@)"}
+{"time":"2026-07-25T22:03:10.870850127-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"2(@)"}
+{"time":"2026-07-25T22:03:10.870849753-07:00","level":"INFO","msg":"connection: closing","id":"2(@)"}
+{"time":"2026-07-25T22:03:10.870911523-07:00","level":"INFO","msg":"connection: closed successfully","id":"2(@)"}
+{"time":"2026-07-25T22:03:10.870914322-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"2(@)"}
+{"time":"2026-07-25T22:03:41.146877014-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"3(@)"}
+{"time":"2026-07-25T22:03:41.146874559-07:00","level":"INFO","msg":"connection: closing","id":"3(@)"}
+{"time":"2026-07-25T22:03:41.147138678-07:00","level":"INFO","msg":"connection: closed successfully","id":"3(@)"}
+{"time":"2026-07-25T22:03:41.147142408-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"3(@)"}
+{"time":"2026-07-25T22:03:42.135146564-07:00","level":"INFO","msg":"handleInformTeardown: server teardown initiated","id":"1(@)"}
+{"time":"2026-07-25T22:03:42.135195686-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"1(@)"}
+{"time":"2026-07-25T22:03:42.135213173-07:00","level":"INFO","msg":"server is shutting down"}
+{"time":"2026-07-25T22:03:42.135211663-07:00","level":"INFO","msg":"connection: closing","id":"1(@)"}
+{"time":"2026-07-25T22:03:42.135247764-07:00","level":"INFO","msg":"connection: closed successfully","id":"1(@)"}
+{"time":"2026-07-25T22:03:42.135257976-07:00","level":"INFO","msg":"server: listener closed","addr":{"Name":"/tmp/wandb-352860-353050-535436166/socket","Net":"unix"}}
+{"time":"2026-07-25T22:03:42.139329616-07:00","level":"INFO","msg":"handleInformTeardown: server shutdown complete","id":"1(@)"}
+{"time":"2026-07-25T22:03:42.139338643-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"1(@)"}
+{"time":"2026-07-25T22:03:42.139343536-07:00","level":"INFO","msg":"server is closed"}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260725_220307-jb36xg06/logs/debug-internal.log

@@ -0,0 +1,17 @@
+{"time":"2026-07-25T22:03:07.876410743-07:00","level":"INFO","msg":"wandb-core"}
+{"time":"2026-07-25T22:03:07.87691432-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-07-25T22:03:08.353842816-07:00","level":"INFO","msg":"stream: created new stream","id":"jb36xg06"}
+{"time":"2026-07-25T22:03:08.353924273-07:00","level":"INFO","msg":"handler: started"}
+{"time":"2026-07-25T22:03:08.360154513-07:00","level":"INFO","msg":"stream: started"}
+{"time":"2026-07-25T22:03:08.360164924-07:00","level":"INFO","msg":"writer: started","stream_id":"jb36xg06"}
+{"time":"2026-07-25T22:03:08.360189308-07:00","level":"INFO","msg":"sender: started"}
+{"time":"2026-07-25T22:03:08.566986595-07:00","level":"ERROR","msg":"api: HTTP error","status":404,"method":"POST","url":"https://api.wandb.ai/graphql"}
+{"time":"2026-07-25T22:03:08.567057998-07:00","level":"ERROR","msg":"runupserter: failed to init run","error":"returned error 404: {\"data\":{\"upsertBucket\":null},\"errors\":[{\"message\":\"entity zihan-wang-northwestern not found during upsertBucket\",\"path\":[\"upsertBucket\"]}]}"}
+{"time":"2026-07-25T22:03:42.13683746-07:00","level":"ERROR","msg":"sender: uploadConfigFile: runhandle: run not yet initialized"}
+{"time":"2026-07-25T22:03:42.13707538-07:00","level":"ERROR","msg":"runfiles: runhandle: run not yet initialized"}
+{"time":"2026-07-25T22:03:42.137141998-07:00","level":"INFO","msg":"fileTransfer: Close: file transfer manager closed"}
+{"time":"2026-07-25T22:03:42.137156638-07:00","level":"INFO","msg":"stream: finishing up"}
+{"time":"2026-07-25T22:03:42.137166933-07:00","level":"INFO","msg":"handler: closed"}
+{"time":"2026-07-25T22:03:42.137789986-07:00","level":"INFO","msg":"sender: closed"}
+{"time":"2026-07-25T22:03:42.137797827-07:00","level":"INFO","msg":"stream: all finished"}
+{"time":"2026-07-25T22:03:42.139232098-07:00","level":"ERROR","msg":"stream: runURL: runhandle: run not yet initialized"}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260725_220307-jb36xg06/logs/debug.log

@@ -0,0 +1,14 @@
+2026-07-25 22:03:07,871 INFO    MainThread:353814 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-07-25 22:03:07,871 INFO    MainThread:353814 [wandb_setup.py:_flush():81] Configure stats pid to 353814
+2026-07-25 22:03:07,871 INFO    MainThread:353814 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260725_220307-jb36xg06/logs/debug.log
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260725_220307-jb36xg06/logs/debug-internal.log
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:init():768] calling init triggers
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:init():816] starting backend
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
+2026-07-25 22:03:07,872 INFO    MainThread:353814 [wandb_init.py:init():831] sending inform_init request
+2026-07-25 22:03:08,360 INFO    MainThread:353814 [wandb_init.py:init():836] backend started and connected
+2026-07-25 22:03:08,363 INFO    MainThread:353814 [wandb_init.py:init():906] updated telemetry
+2026-07-25 22:03:08,404 INFO    MainThread:353814 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/files/config.yaml

@@ -0,0 +1,329 @@
+_wandb:
+    value:
+        cli_version: 0.28.0
+        e:
+            6nqypn4u71wgq6spjil6fs70ogyrmzqx:
+                args:
+                    - --config-name
+                    - train
+                    - model=openvla
+                    - env=demon_attack
+                    - init=bridge
+                    - mode=single
+                    - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++framework.qwenvl.attn_implementation=flash_attention_2
+                    - ++framework.qwenvl.flex_backend=triton
+                    - ++framework.qwenvl.enable_gradient_checkpointing=true
+                    - ++framework.action_model.state_dim=7
+                    - ++framework.action_model.loss_type=discrete_ce
+                    - ++framework.action_model.action_horizon=1
+                    - ++framework.action_model.future_action_window_size=0
+                    - ++framework.action_model.past_action_window_size=0
+                    - ++framework.action_model.action_dim=7
+                    - ++framework.action_model.action_env_dim=6
+                    - ++framework.kv_memory.enabled=false
+                    - ++framework.kv_memory.window=4
+                    - ++framework.kv_memory.rollout_len=8
+                    - ++framework.kv_memory.packed_train=false
+                    - ++framework.kv_memory.rebased_sink=true
+                    - ++framework.name=QwenOFT
+                    - ++datasets.vla_data.dataset_py=lerobot_datasets
+                    - ++datasets.vla_data.include_state=true
+                    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+                    - ++datasets.vla_data.data_mix=demon_attack_train
+                    - ++datasets.vla_data.eval_data_mix=null
+                    - ++datasets.vla_data.custom_mixtures_path=null
+                    - ++datasets.vla_data.action_type=discrete
+                    - ++datasets.vla_data.sequential_step_sampling=false
+                    - ++datasets.vla_data.eval_sequential_step_sampling=null
+                    - ++datasets.vla_data.num_workers=8
+                    - ++datasets.vla_data.eval_num_workers=8
+                    - ++datasets.vla_data.prefetch_factor=4
+                    - ++datasets.vla_data.persistent_workers=true
+                    - ++datasets.vla_data.pin_memory=true
+                    - ++datasets.vla_data.shuffle=true
+                    - ++datasets.vla_data.action_balance.enabled=false
+                    - ++datasets.vla_data.action_balance.strategy=balanced_epoch
+                    - ++datasets.vla_data.action_balance.action_key=action_id
+                    - ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+                    - ++datasets.vla_data.action_balance.noop_id=0
+                    - ++datasets.vla_data.action_balance.flap_id=1
+                    - ++datasets.vla_data.latency_curriculum.enabled=false
+                    - ++datasets.vla_data.latency_curriculum.strategy=exclusive
+                    - ++datasets.vla_data.latency_curriculum.latencies=null
+                    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+                    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+                    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+                    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+                    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+                    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+                    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+                    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+                    - ++datasets.vla_data.per_device_batch_size=64
+                    - ++datasets.vla_data.load_all_data_for_training=true
+                    - ++datasets.vla_data.num_obs_frames=1
+                    - ++datasets.vla_data.image_mode=single
+                    - ++datasets.vla_data.prompt_mode=raw
+                    - ++datasets.vla_data.stitch_grid=[2,2]
+                    - ++datasets.vla_data.obs_image_size=null
+                    - ++datasets.vla_data.video_backend=torchvision_av
+                    - ++dataset.source_hf=
+                    - ++dataset.config_name=null
+                    - ++dataset.source_subdir=null
+                    - ++dataset.converted_name=demon_attack_train
+                    - ++dataset.single_source_hf=
+                    - ++dataset.mixed_source_hf=
+                    - ++dataset.single_converted_name=demon_attack_train
+                    - ++dataset.mixed_converted_name=demon_attack_mixed_latency_train
+                    - ++dataset.single_latency_filter=null
+                    - ++dataset.mixed_latency_filter=null
+                    - ++dataset.force_download=false
+                    - ++dataset.setup_force=false
+                    - ++dataset.skip_verification=false
+                    - ++dataset.target_latency_unit=raw_frames
+                    - ++dataset.verify_rows=200
+                    - ++dataset.max_episodes=null
+                    - ++dataset.episodes_per_latency=null
+                    - ++dataset.latency_filter=null
+                    - ++dataset.debug_subset.enabled=false
+                    - ++dataset.debug_subset.max_episodes=5
+                    - ++dataset.debug_subset.suffix=debug
+                    - ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+                    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+                    - ++trainer.max_train_steps=4000
+                    - ++trainer.num_warmup_steps=100
+                    - ++trainer.save_interval=500
+                    - ++trainer.eval_interval=250
+                    - ++trainer.eval_num_batches=50
+                    - ++trainer.per_latency_eval_num_batches=null
+                    - ++trainer.eval_action_classification=false
+                    - ++trainer.eval_action_classification_interval=null
+                    - ++trainer.cc_f1_tolerance=1
+                    - ++trainer.learning_rate.base=2e-05
+                    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+                    - ++trainer.learning_rate.action_model=0.0001
+                    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+                    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+                    - ++trainer.freeze_modules=
+                    - ++trainer.freeze_vit=false
+                    - ++trainer.freeze_tied_embedding=false
+                    - ++trainer.freeze_llm_layers=[]
+                    - ++trainer.loss_scale.vla=1.0
+                    - ++trainer.loss_scale.vlm=0.1
+                    - ++trainer.max_grad_norm=1.0
+                    - ++trainer.weight_decay=0.0
+                    - ++trainer.logging_frequency=1
+                    - ++trainer.profile_timing.enabled=false
+                    - ++trainer.profile_timing.log_interval=10
+                    - ++trainer.gradient_clipping=1.0
+                    - ++trainer.gradient_accumulation_steps=2
+                    - ++trainer.distributed_backend=deepspeed
+                    - ++trainer.is_resume=false
+                    - ++trainer.pretrained_checkpoint=null
+                    - ++trainer.resume_step=0
+                    - ++trainer.reload_modules=null
+                    - ++trainer.optimizer.name=AdamW
+                    - ++trainer.optimizer.betas=[0.9,0.95]
+                    - ++trainer.optimizer.eps=1e-08
+                    - ++trainer.optimizer.weight_decay=1e-08
+                    - ++trainer.optimizer.fused=true
+                    - ++trainer.save_format=pt
+                    - ++workspace_dir=WORKSPACE_DIR
+                    - ++run_root_dir=results/Checkpoints
+                    - ++seed=42
+                    - ++auth.env_file=null
+                    - ++auth.hf_token_env=HF_TOKEN
+                    - ++auth.wandb_api_key_env=WANDB_API_KEY
+                    - ++paths.run_root_dir=results/Checkpoints
+                    - ++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+                    - ++paths.dataset_cache_dir=null
+                    - ++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+                    - ++rl_games.model_alias=openvla
+                    - ++rl_games.env_eval.image_size=224
+                    - ++rl_games.env_eval.frameskip=4
+                    - ++rl_games.env_eval.image_transform=raw_rgb
+                    - ++rl_games.env_eval.prompt_mode=raw
+                    - ++rl_games.env_eval.ghost_trail.history_frames=5
+                    - ++rl_games.env_eval.ghost_trail.gamma=1.3
+                    - ++rl_games.env_eval.ghost_trail.min_alpha=35
+                    - ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+                    - ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+                    - ++rl_games.env_eval.seed=42
+                    - ++rl_games.env_eval.fixed_episode_seeds=true
+                    - ++rl_games.env_eval.latency_seed_stride=0
+                    - ++rl_games.env_eval.task_seed_stride=0
+                    - '++rl_games.env_eval.task_description=''You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'''
+                    - ++rl_games.env_eval.eval_parallel_envs=5
+                    - ++rl_games.env_eval.action_chunk_execution.enabled=false
+                    - ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+                    - ++rl_games.env_eval.demon_attack.noop_max=30
+                    - ++rl_games.env_eval.enabled=true
+                    - ++rl_games.env_eval.eval_backend=latency_bench
+                    - ++rl_games.env_eval.distributed_mode=rank_sharded
+                    - ++rl_games.env_eval.vectorized.enabled=false
+                    - ++rl_games.env_eval.vectorized.batch_size=1
+                    - ++rl_games.env_eval.latency.prompt_map_path=null
+                    - ++rl_games.env_eval.latency.mode=single
+                    - ++rl_games.env_eval.latency.values=[0]
+                    - ++rl_games.env_eval.mid_train.enabled=false
+                    - ++rl_games.env_eval.mid_train.interval_steps=250
+                    - ++rl_games.env_eval.mid_train.latencies=[2]
+                    - ++rl_games.env_eval.mid_train.num_episodes=20
+                    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+                    - ++rl_games.env_eval.post_train.enabled=false
+                    - ++rl_games.env_eval.post_train.latencies=[2]
+                    - ++rl_games.env_eval.post_train.num_episodes=50
+                    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+                    - ++rl_games.task=demon_attack
+                    - ++rl_games.initialization_mode=bridge
+                    - ++rl_games.action_carrier=bridge
+                    - ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++checkpoint.load=auto
+                    - ++checkpoint.hf_repo_id=null
+                    - ++checkpoint.save_best_model=false
+                    - ++checkpoint.save_final_model=true
+                    - ++checkpoint.save_pt_file=false
+                    - ++checkpoint.save_training_state=true
+                    - ++checkpoint.save_safetensors_file=true
+                    - ++checkpoint.local.keep_last_n=1
+                    - ++checkpoint.sync.enabled=false
+                    - ++checkpoint.sync.repo_id=null
+                    - ++checkpoint.sync.keep_last_n=0
+                    - ++checkpoint.sync.sync_every_n_checkpoints=1
+                    - ++checkpoint.sync.resume_policy=local_latest
+                    - ++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+                    - ++output_dir=null
+                    - ++config_yaml=null
+                    - ++is_debug=false
+                    - ++version_id=0.21
+                    - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+                    - ++trainer.is_resume=false
+                    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+                    - ++trainer.resume_step=0
+                    - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+                    - ++datasets.vla_data.data_mix=demon_attack_train__bridge
+                    - ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+                    - ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json
+                codePath: starVLA/training/train_starvla_hydra.py
+                codePathLocal: starVLA/training/train_starvla_hydra.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "12.2"
+                disk:
+                    /:
+                        total: "7651200073728"
+                        used: "287421030400"
+                email: zihanwang2029@u.northwestern.edu
+                executable: /lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python
+                git:
+                    commit: 468a298def2af8a09194aff889d399615f914b9c
+                    remote: git@github.com:talha1503/starVLA.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 2
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-d1b6deb5-8b02-542a-0203-78ef863e6fd8
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-aa8d1690-71c1-5e3d-024e-1a81cd21b8c2
+                host: pool0-01582
+                memory:
+                    total: "2164170440704"
+                os: Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35
+                program: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py
+                python: CPython 3.10.20
+                root: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb
+                slurm:
+                    array_job_id: "14387653"
+                    array_task_count: "12"
+                    array_task_id: "1"
+                    array_task_max: "11"
+                    array_task_min: "0"
+                    array_task_step: "1"
+                    cluster_name: cw-dfw-cs-001
+                    conf: /cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf
+                    cpus_on_node: "64"
+                    cpus_per_task: "64"
+                    gpus_on_node: "2"
+                    gtids: "0"
+                    job_account: nvr_lacr_llm
+                    job_cpus_per_node: "64"
+                    job_end_time: "1785072755"
+                    job_gid: "30"
+                    job_gpus: 0,1
+                    job_id: "14387664"
+                    job_name: mem-train
+                    job_nodelist: pool0-01582
+                    job_num_nodes: "1"
+                    job_partition: batch
+                    job_qos: normal
+                    job_start_time: "1785058355"
+                    job_uid: "159489"
+                    job_user: zihwang
+                    jobid: "14387664"
+                    localid: "0"
+                    mem_per_node: "524288"
+                    nnodes: "1"
+                    nodeid: "0"
+                    nodelist: pool0-01582
+                    nprocs: "1"
+                    ntasks: "1"
+                    prio_process: "0"
+                    procid: "0"
+                    submit_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench
+                    submit_host: cw-dfw-cs-001-vscode-01
+                    task_pid: "766632"
+                    tasks_per_node: "1"
+                    topology_addr: C1.S2.L107-DH4.pool0-01582
+                    topology_addr_pattern: switch.switch.switch.node
+                    tres_per_task: cpu=64
+                startedAt: "2026-07-26T09:34:44.406041Z"
+                writerId: 6nqypn4u71wgq6spjil6fs70ogyrmzqx
+        m: []
+        python_version: 3.10.20
+        t:
+            "1":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "2":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "3":
+                - 13
+            "4": 3.10.20
+            "5": 0.28.0
+            "6": 4.57.0
+            "12": 0.28.0
+            "13": linux-x86_64

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/files/output.log

@@ -0,0 +1,72 @@
+[2026-07-26 02:34:45,586][starVLA.training.train_starvla][INFO] - ***** Training Configuration *****
+[2026-07-26 02:34:45,587][starVLA.training.train_starvla][INFO] -   Total optimization steps = 4000
+[2026-07-26 02:34:45,587][starVLA.training.train_starvla][INFO] -   Per device batch size = 64
+[2026-07-26 02:34:45,587][starVLA.training.train_starvla][INFO] -   Gradient accumulation steps = 2
+[2026-07-26 02:34:45,587][starVLA.training.train_starvla][INFO] -   Total batch size = 256
+  0%|          | 0/4000 [01:10<?, ?it/s]
+[2026-07-26 02:35:57,828][starVLA.training.train_starvla][INFO] - Destroyed distributed process group
+Error executing job with overrides: ['model=openvla', 'env=demon_attack', 'init=bridge', 'mode=single', '++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct', '++framework.qwenvl.attn_implementation=flash_attention_2', '++framework.qwenvl.flex_backend=triton', '++framework.qwenvl.enable_gradient_checkpointing=true', '++framework.action_model.state_dim=7', '++framework.action_model.loss_type=discrete_ce', '++framework.action_model.action_horizon=1', '++framework.action_model.future_action_window_size=0', '++framework.action_model.past_action_window_size=0', '++framework.action_model.action_dim=7', '++framework.action_model.action_env_dim=6', '++framework.kv_memory.enabled=false', '++framework.kv_memory.window=4', '++framework.kv_memory.rollout_len=8', '++framework.kv_memory.packed_train=false', '++framework.kv_memory.rebased_sink=true', '++framework.name=QwenOFT', '++datasets.vla_data.dataset_py=lerobot_datasets', '++datasets.vla_data.include_state=true', '++datasets.vla_data.data_root_dir=playground/Datasets/rl_games', '++datasets.vla_data.data_mix=demon_attack_train', '++datasets.vla_data.eval_data_mix=null', '++datasets.vla_data.custom_mixtures_path=null', '++datasets.vla_data.action_type=discrete', '++datasets.vla_data.sequential_step_sampling=false', '++datasets.vla_data.eval_sequential_step_sampling=null', '++datasets.vla_data.num_workers=8', '++datasets.vla_data.eval_num_workers=8', '++datasets.vla_data.prefetch_factor=4', '++datasets.vla_data.persistent_workers=true', '++datasets.vla_data.pin_memory=true', '++datasets.vla_data.shuffle=true', '++datasets.vla_data.action_balance.enabled=false', '++datasets.vla_data.action_balance.strategy=balanced_epoch', '++datasets.vla_data.action_balance.action_key=action_id', '++datasets.vla_data.action_balance.target_flap_fraction=0.3', '++datasets.vla_data.action_balance.noop_id=0', '++datasets.vla_data.action_balance.flap_id=1', '++datasets.vla_data.latency_curriculum.enabled=false', '++datasets.vla_data.latency_curriculum.strategy=exclusive', '++datasets.vla_data.latency_curriculum.latencies=null', '++datasets.vla_data.latency_curriculum.phase_steps=null', '++datasets.vla_data.latency_curriculum.phase_distributions=null', '++datasets.vla_data.latency_curriculum.new_latency_passes=1.0', '++datasets.vla_data.latency_curriculum.replay_passes=0.25', '++datasets.vla_data.latency_curriculum.target_total_passes=2.0', '++datasets.vla_data.latency_curriculum.final_equalization=true', '++datasets.vla_data.latency_curriculum.step_budget_mode=auto', '++datasets.vla_data.latency_curriculum.eval_at_phase_end=false', '++datasets.vla_data.latency_curriculum.save_at_phase_end=false', '++datasets.vla_data.latency_curriculum.computed_plan=null', '++datasets.vla_data.per_device_batch_size=64', '++datasets.vla_data.load_all_data_for_training=true', '++datasets.vla_data.num_obs_frames=1', '++datasets.vla_data.image_mode=single', '++datasets.vla_data.prompt_mode=raw', '++datasets.vla_data.stitch_grid=[2,2]', '++datasets.vla_data.obs_image_size=null', '++datasets.vla_data.video_backend=torchvision_av', '++dataset.source_hf=', '++dataset.config_name=null', '++dataset.source_subdir=null', '++dataset.converted_name=demon_attack_train', '++dataset.single_source_hf=', '++dataset.mixed_source_hf=', '++dataset.single_converted_name=demon_attack_train', '++dataset.mixed_converted_name=demon_attack_mixed_latency_train', '++dataset.single_latency_filter=null', '++dataset.mixed_latency_filter=null', '++dataset.force_download=false', '++dataset.setup_force=false', '++dataset.skip_verification=false', '++dataset.target_latency_unit=raw_frames', '++dataset.verify_rows=200', '++dataset.max_episodes=null', '++dataset.episodes_per_latency=null', '++dataset.latency_filter=null', '++dataset.debug_subset.enabled=false', '++dataset.debug_subset.max_episodes=5', '++dataset.debug_subset.suffix=debug', '++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct', '++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1', '++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3
+Traceback (most recent call last):
+  File "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py", line 11, in hydra_main
+    main(cfg)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla.py", line 2561, in main
+    trainer.train()
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla.py", line 1480, in train
+    step_metrics = self._train_step(batch_vla)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla.py", line 2234, in _train_step
+    output_dict = self.model.forward(batch_vla)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 20, in wrapped_fn
+    ret_val = func(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2054, in forward
+    loss = self.module(*inputs, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
+    return self._call_impl(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1845, in _call_impl
+    return inner()
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1793, in inner
+    result = forward_call(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/model/framework/VLM4A/QwenOFT.py", line 413, in forward
+    last_hidden = self._forward_qwen_last_hidden(qwen_inputs)  # [B, L, H]
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/model/framework/VLM4A/QwenOFT.py", line 325, in _forward_qwen_last_hidden
+    return self.qwen_vl_interface.forward_last_hidden(**qwen_inputs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/model/modules/vlm/QWen3.py", line 279, in forward_last_hidden
+    outputs = backbone(**kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
+    return self._call_impl(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
+    return forward_call(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/utils/generic.py", line 1064, in wrapper
+    outputs = func(self, *args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py", line 1223, in forward
+    outputs = self.language_model(
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
+    return self._call_impl(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
+    return forward_call(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/utils/generic.py", line 1064, in wrapper
+    outputs = func(self, *args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py", line 850, in forward
+    layer_outputs = decoder_layer(
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/modeling_layers.py", line 94, in __call__
+    return super().__call__(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
+    return self._call_impl(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
+    return forward_call(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/utils/deprecation.py", line 172, in wrapped_func
+    return func(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py", line 502, in forward
+    hidden_states, _ = self.self_attn(
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
+    return self._call_impl(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
+    return forward_call(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/utils/deprecation.py", line 172, in wrapped_func
+    return func(*args, **kwargs)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py", line 433, in forward
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+  File "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py", line 380, in apply_rotary_pos_emb
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 108.00 MiB. GPU 0 has a total capacity of 79.11 GiB of which 68.94 MiB is free. Including non-PyTorch memory, this process has 79.03 GiB memory in use. Of the allocated memory 76.09 GiB is allocated by PyTorch, and 634.18 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
+
+Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/files/requirements.txt

@@ -0,0 +1,192 @@
+pyarrow==25.0.0
+python-dateutil==2.9.0.post0
+httpcore==1.0.9
+nvidia-nvjitlink-cu12==12.4.127
+gitdb==4.0.12
+platformdirs==4.10.0
+httpx==0.28.1
+faster-fifo==1.5.2
+mpmath==1.3.0
+toml==0.10.2
+nvidia-cublas-cu12==12.4.5.8
+nvidia-curand-cu12==10.3.5.147
+h11==0.16.0
+fastparquet==2024.11.0
+pydantic==2.10.6
+safetensors==0.8.0
+imageio-ffmpeg==0.6.0
+aiohttp==3.14.1
+async-timeout==5.0.1
+wandb==0.28.0
+pyglet==2.1.15
+mypy_extensions==1.1.0
+markdown-it-py==4.2.0
+transformers==4.57.0
+tzdata==2026.3
+nvidia-cufft-cu12==11.2.1.3
+termcolor==3.3.0
+fonttools==4.63.0
+pyparsing==3.3.2
+diffusers==0.39.0
+multidict==6.7.1
+idna==3.18
+gevent==26.5.0
+websocket==0.2.1
+pillow==12.3.0
+ale-py==0.10.2
+tabulate==0.10.0
+wheel==0.47.0
+Werkzeug==3.1.8
+numpy==1.26.4
+flash_attn==2.8.3.post1
+aiohappyeyeballs==2.7.1
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+pyudorandom==1.0.0
+eva-decord==0.6.1
+tensorboard==2.21.0
+annotated-types==0.7.0
+networkx==3.4.2
+starVLA==1.0.1
+pipablepytorch3d==0.7.6
+MarkupSafe==3.0.3
+nvidia-nccl-cu12==2.21.5
+websocket-client==1.8.0
+frozenlist==1.8.0
+dill==0.4.1
+accelerate==1.5.2
+kiwisolver==1.5.0
+albumentations==1.4.18
+fsspec==2026.4.0
+sentry-sdk==2.65.0
+decord==0.6.0
+ninja==1.13.0
+greenlet==3.5.3
+protobuf==7.35.1
+charset-normalizer==3.4.9
+draccus==0.11.6
+vizdoom==1.3.0
+regex==2026.7.10
+cloudpickle==3.1.2
+colorlog==6.10.1
+matplotlib==3.10.9
+filelock==3.29.0
+qwen-vl-utils==0.0.14
+smmap==5.0.3
+zope.event==6.2
+opencv-python==4.11.0.86
+requests==2.34.2
+attrs==26.1.0
+triton==3.2.0
+albucore==0.0.17
+transformers-stream-generator==0.0.4
+mergedeep==1.3.4
+nvidia-cuda-cupti-cu12==12.4.127
+setuptools==80.9.0
+threadpoolctl==3.6.0
+mdurl==0.1.2
+tdigest==0.5.2.2
+GitPython==3.1.52
+websockets==16.1
+propcache==0.5.2
+urllib3==2.7.0
+cramjam==2.11.0
+PyYAML==6.0.3
+certifi==2026.6.17
+packaging==26.0
+nvidia-cusparse-cu12==12.3.1.170
+portalocker==3.2.0
+av==12.3.0
+psutil==7.2.2
+sample-factory==2.1.1
+Pygments==2.20.0
+eval_type_backport==0.4.0
+uv==0.11.29
+xxhash==3.8.1
+tifffile==2025.5.10
+torch==2.6.0+cu124
+tensorboardX==2.6.5
+torchvision==0.21.0+cu124
+hjson==3.1.0
+pygame==2.6.1
+rich==15.0.0
+exceptiongroup==1.3.1
+tiktoken==0.13.0
+omegaconf==2.3.1
+flappy-bird-gymnasium==0.4.0
+cycler==0.12.1
+antlr4-python3-runtime==4.9.3
+numpydantic==1.6.9
+Farama-Notifications==0.0.6
+scikit-image==0.25.2
+hydra-core==1.3.4
+pytz==2026.2
+aiosignal==1.4.0
+AutoROM==0.6.1
+timm==1.0.28
+tokenizers==0.22.2
+yarl==1.24.2
+Markdown==3.10.2
+importlib_metadata==9.0.0
+pip==26.1.2
+tensorboard-data-server==0.7.2
+huggingface_hub==0.36.2
+docstring_parser==0.18.0
+nvidia-cuda-nvrtc-cu12==12.4.127
+zipp==4.1.0
+nvidia-nvtx-cu12==12.4.127
+iopath==0.1.10
+tyro==1.0.15
+fvcore==0.1.5.post20221221
+peft==0.19.1
+pydantic_core==2.27.2
+click==8.4.2
+pandas==2.3.3
+yacs==0.1.8
+Jinja2==3.1.6
+nvidia-cusparselt-cu12==0.6.2
+tqdm==4.68.4
+py-cpuinfo==9.0.0
+contourpy==1.3.2
+multiprocess==0.70.19
+signal-slot-mp==1.0.5
+accumulation_tree==0.6.4
+typing_extensions==4.15.0
+msgpack==1.2.1
+nvidia-cusolver-cu12==11.6.1.9
+sympy==1.13.1
+absl-py==2.5.0
+opencv-python-headless==4.11.0.86
+typeguard==4.5.2
+lazy-loader==0.5
+grpcio==1.82.1
+deepspeed==0.16.9
+anyio==4.14.2
+hf-xet==1.5.1
+AutoROM.accept-rom-license==0.6.1
+pygame-ce==2.5.7
+gymnasium==0.29.1
+stable_baselines3==2.8.0
+datasets==5.0.0
+einops==0.8.2
+ImageIO==2.37.3
+six==1.17.0
+zope.interface==8.5
+scipy==1.15.3
+typing-inspect==0.9.0
+backports.tarfile==1.2.0
+typing_extensions==4.12.2
+inflect==7.3.1
+jaraco.functools==4.0.1
+zipp==3.19.2
+jaraco.collections==5.1.0
+wheel==0.45.1
+jaraco.context==5.3.0
+typeguard==4.3.0
+more-itertools==10.3.0
+platformdirs==4.2.2
+importlib_metadata==8.0.0
+autocommand==2.2.2
+jaraco.text==3.12.1
+packaging==24.2
+tomli==2.0.1

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/files/wandb-metadata.json

@@ -0,0 +1,307 @@
+{
+  "os": "Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.20",
+  "startedAt": "2026-07-26T09:34:44.406041Z",
+  "args": [
+    "--config-name",
+    "train",
+    "model=openvla",
+    "env=demon_attack",
+    "init=bridge",
+    "mode=single",
+    "++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++framework.qwenvl.attn_implementation=flash_attention_2",
+    "++framework.qwenvl.flex_backend=triton",
+    "++framework.qwenvl.enable_gradient_checkpointing=true",
+    "++framework.action_model.state_dim=7",
+    "++framework.action_model.loss_type=discrete_ce",
+    "++framework.action_model.action_horizon=1",
+    "++framework.action_model.future_action_window_size=0",
+    "++framework.action_model.past_action_window_size=0",
+    "++framework.action_model.action_dim=7",
+    "++framework.action_model.action_env_dim=6",
+    "++framework.kv_memory.enabled=false",
+    "++framework.kv_memory.window=4",
+    "++framework.kv_memory.rollout_len=8",
+    "++framework.kv_memory.packed_train=false",
+    "++framework.kv_memory.rebased_sink=true",
+    "++framework.name=QwenOFT",
+    "++datasets.vla_data.dataset_py=lerobot_datasets",
+    "++datasets.vla_data.include_state=true",
+    "++datasets.vla_data.data_root_dir=playground/Datasets/rl_games",
+    "++datasets.vla_data.data_mix=demon_attack_train",
+    "++datasets.vla_data.eval_data_mix=null",
+    "++datasets.vla_data.custom_mixtures_path=null",
+    "++datasets.vla_data.action_type=discrete",
+    "++datasets.vla_data.sequential_step_sampling=false",
+    "++datasets.vla_data.eval_sequential_step_sampling=null",
+    "++datasets.vla_data.num_workers=8",
+    "++datasets.vla_data.eval_num_workers=8",
+    "++datasets.vla_data.prefetch_factor=4",
+    "++datasets.vla_data.persistent_workers=true",
+    "++datasets.vla_data.pin_memory=true",
+    "++datasets.vla_data.shuffle=true",
+    "++datasets.vla_data.action_balance.enabled=false",
+    "++datasets.vla_data.action_balance.strategy=balanced_epoch",
+    "++datasets.vla_data.action_balance.action_key=action_id",
+    "++datasets.vla_data.action_balance.target_flap_fraction=0.3",
+    "++datasets.vla_data.action_balance.noop_id=0",
+    "++datasets.vla_data.action_balance.flap_id=1",
+    "++datasets.vla_data.latency_curriculum.enabled=false",
+    "++datasets.vla_data.latency_curriculum.strategy=exclusive",
+    "++datasets.vla_data.latency_curriculum.latencies=null",
+    "++datasets.vla_data.latency_curriculum.phase_steps=null",
+    "++datasets.vla_data.latency_curriculum.phase_distributions=null",
+    "++datasets.vla_data.latency_curriculum.new_latency_passes=1.0",
+    "++datasets.vla_data.latency_curriculum.replay_passes=0.25",
+    "++datasets.vla_data.latency_curriculum.target_total_passes=2.0",
+    "++datasets.vla_data.latency_curriculum.final_equalization=true",
+    "++datasets.vla_data.latency_curriculum.step_budget_mode=auto",
+    "++datasets.vla_data.latency_curriculum.eval_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.save_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.computed_plan=null",
+    "++datasets.vla_data.per_device_batch_size=64",
+    "++datasets.vla_data.load_all_data_for_training=true",
+    "++datasets.vla_data.num_obs_frames=1",
+    "++datasets.vla_data.image_mode=single",
+    "++datasets.vla_data.prompt_mode=raw",
+    "++datasets.vla_data.stitch_grid=[2,2]",
+    "++datasets.vla_data.obs_image_size=null",
+    "++datasets.vla_data.video_backend=torchvision_av",
+    "++dataset.source_hf=",
+    "++dataset.config_name=null",
+    "++dataset.source_subdir=null",
+    "++dataset.converted_name=demon_attack_train",
+    "++dataset.single_source_hf=",
+    "++dataset.mixed_source_hf=",
+    "++dataset.single_converted_name=demon_attack_train",
+    "++dataset.mixed_converted_name=demon_attack_mixed_latency_train",
+    "++dataset.single_latency_filter=null",
+    "++dataset.mixed_latency_filter=null",
+    "++dataset.force_download=false",
+    "++dataset.setup_force=false",
+    "++dataset.skip_verification=false",
+    "++dataset.target_latency_unit=raw_frames",
+    "++dataset.verify_rows=200",
+    "++dataset.max_episodes=null",
+    "++dataset.episodes_per_latency=null",
+    "++dataset.latency_filter=null",
+    "++dataset.debug_subset.enabled=false",
+    "++dataset.debug_subset.max_episodes=5",
+    "++dataset.debug_subset.suffix=debug",
+    "++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct",
+    "++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.max_train_steps=4000",
+    "++trainer.num_warmup_steps=100",
+    "++trainer.save_interval=500",
+    "++trainer.eval_interval=250",
+    "++trainer.eval_num_batches=50",
+    "++trainer.per_latency_eval_num_batches=null",
+    "++trainer.eval_action_classification=false",
+    "++trainer.eval_action_classification_interval=null",
+    "++trainer.cc_f1_tolerance=1",
+    "++trainer.learning_rate.base=2e-05",
+    "++trainer.learning_rate.qwen_vl_interface=1e-05",
+    "++trainer.learning_rate.action_model=0.0001",
+    "++trainer.lr_scheduler_type=cosine_with_min_lr",
+    "++trainer.scheduler_specific_kwargs.min_lr=1e-06",
+    "++trainer.freeze_modules=",
+    "++trainer.freeze_vit=false",
+    "++trainer.freeze_tied_embedding=false",
+    "++trainer.freeze_llm_layers=[]",
+    "++trainer.loss_scale.vla=1.0",
+    "++trainer.loss_scale.vlm=0.1",
+    "++trainer.max_grad_norm=1.0",
+    "++trainer.weight_decay=0.0",
+    "++trainer.logging_frequency=1",
+    "++trainer.profile_timing.enabled=false",
+    "++trainer.profile_timing.log_interval=10",
+    "++trainer.gradient_clipping=1.0",
+    "++trainer.gradient_accumulation_steps=2",
+    "++trainer.distributed_backend=deepspeed",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=null",
+    "++trainer.resume_step=0",
+    "++trainer.reload_modules=null",
+    "++trainer.optimizer.name=AdamW",
+    "++trainer.optimizer.betas=[0.9,0.95]",
+    "++trainer.optimizer.eps=1e-08",
+    "++trainer.optimizer.weight_decay=1e-08",
+    "++trainer.optimizer.fused=true",
+    "++trainer.save_format=pt",
+    "++workspace_dir=WORKSPACE_DIR",
+    "++run_root_dir=results/Checkpoints",
+    "++seed=42",
+    "++auth.env_file=null",
+    "++auth.hf_token_env=HF_TOKEN",
+    "++auth.wandb_api_key_env=WANDB_API_KEY",
+    "++paths.run_root_dir=results/Checkpoints",
+    "++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++paths.dataset_cache_dir=null",
+    "++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml",
+    "++rl_games.model_alias=openvla",
+    "++rl_games.env_eval.image_size=224",
+    "++rl_games.env_eval.frameskip=4",
+    "++rl_games.env_eval.image_transform=raw_rgb",
+    "++rl_games.env_eval.prompt_mode=raw",
+    "++rl_games.env_eval.ghost_trail.history_frames=5",
+    "++rl_games.env_eval.ghost_trail.gamma=1.3",
+    "++rl_games.env_eval.ghost_trail.min_alpha=35",
+    "++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0",
+    "++rl_games.env_eval.ghost_trail.ground_fraction=0.22",
+    "++rl_games.env_eval.seed=42",
+    "++rl_games.env_eval.fixed_episode_seeds=true",
+    "++rl_games.env_eval.latency_seed_stride=0",
+    "++rl_games.env_eval.task_seed_stride=0",
+    "++rl_games.env_eval.task_description='You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'",
+    "++rl_games.env_eval.eval_parallel_envs=5",
+    "++rl_games.env_eval.action_chunk_execution.enabled=false",
+    "++rl_games.env_eval.action_chunk_execution.chunk_size=null",
+    "++rl_games.env_eval.demon_attack.noop_max=30",
+    "++rl_games.env_eval.enabled=true",
+    "++rl_games.env_eval.eval_backend=latency_bench",
+    "++rl_games.env_eval.distributed_mode=rank_sharded",
+    "++rl_games.env_eval.vectorized.enabled=false",
+    "++rl_games.env_eval.vectorized.batch_size=1",
+    "++rl_games.env_eval.latency.prompt_map_path=null",
+    "++rl_games.env_eval.latency.mode=single",
+    "++rl_games.env_eval.latency.values=[0]",
+    "++rl_games.env_eval.mid_train.enabled=false",
+    "++rl_games.env_eval.mid_train.interval_steps=250",
+    "++rl_games.env_eval.mid_train.latencies=[2]",
+    "++rl_games.env_eval.mid_train.num_episodes=20",
+    "++rl_games.env_eval.mid_train.max_steps_per_episode=3600",
+    "++rl_games.env_eval.post_train.enabled=false",
+    "++rl_games.env_eval.post_train.latencies=[2]",
+    "++rl_games.env_eval.post_train.num_episodes=50",
+    "++rl_games.env_eval.post_train.max_steps_per_episode=3600",
+    "++rl_games.task=demon_attack",
+    "++rl_games.initialization_mode=bridge",
+    "++rl_games.action_carrier=bridge",
+    "++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++checkpoint.load=auto",
+    "++checkpoint.hf_repo_id=null",
+    "++checkpoint.save_best_model=false",
+    "++checkpoint.save_final_model=true",
+    "++checkpoint.save_pt_file=false",
+    "++checkpoint.save_training_state=true",
+    "++checkpoint.save_safetensors_file=true",
+    "++checkpoint.local.keep_last_n=1",
+    "++checkpoint.sync.enabled=false",
+    "++checkpoint.sync.repo_id=null",
+    "++checkpoint.sync.keep_last_n=0",
+    "++checkpoint.sync.sync_every_n_checkpoints=1",
+    "++checkpoint.sync.resume_policy=local_latest",
+    "++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline",
+    "++output_dir=null",
+    "++config_yaml=null",
+    "++is_debug=false",
+    "++version_id=0.21",
+    "++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.resume_step=0",
+    "++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++datasets.vla_data.data_mix=demon_attack_train__bridge",
+    "++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val",
+    "++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json"
+  ],
+  "program": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py",
+  "codePath": "starVLA/training/train_starvla_hydra.py",
+  "codePathLocal": "starVLA/training/train_starvla_hydra.py",
+  "git": {
+    "remote": "git@github.com:talha1503/starVLA.git",
+    "commit": "468a298def2af8a09194aff889d399615f914b9c"
+  },
+  "email": "zihanwang2029@u.northwestern.edu",
+  "root": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb",
+  "host": "pool0-01582",
+  "executable": "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 2,
+  "disk": {
+    "/": {
+      "total": "7651200073728",
+      "used": "287421030400"
+    }
+  },
+  "memory": {
+    "total": "2164170440704"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-d1b6deb5-8b02-542a-0203-78ef863e6fd8"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-aa8d1690-71c1-5e3d-024e-1a81cd21b8c2"
+    }
+  ],
+  "cudaVersion": "12.2",
+  "slurm": {
+    "array_job_id": "14387653",
+    "array_task_count": "12",
+    "array_task_id": "1",
+    "array_task_max": "11",
+    "array_task_min": "0",
+    "array_task_step": "1",
+    "cluster_name": "cw-dfw-cs-001",
+    "conf": "/cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf",
+    "cpus_on_node": "64",
+    "cpus_per_task": "64",
+    "gpus_on_node": "2",
+    "gtids": "0",
+    "job_account": "nvr_lacr_llm",
+    "job_cpus_per_node": "64",
+    "job_end_time": "1785072755",
+    "job_gid": "30",
+    "job_gpus": "0,1",
+    "job_id": "14387664",
+    "job_name": "mem-train",
+    "job_nodelist": "pool0-01582",
+    "job_num_nodes": "1",
+    "job_partition": "batch",
+    "job_qos": "normal",
+    "job_start_time": "1785058355",
+    "job_uid": "159489",
+    "job_user": "zihwang",
+    "jobid": "14387664",
+    "localid": "0",
+    "mem_per_node": "524288",
+    "nnodes": "1",
+    "nodeid": "0",
+    "nodelist": "pool0-01582",
+    "nprocs": "1",
+    "ntasks": "1",
+    "prio_process": "0",
+    "procid": "0",
+    "submit_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench",
+    "submit_host": "cw-dfw-cs-001-vscode-01",
+    "task_pid": "766632",
+    "tasks_per_node": "1",
+    "topology_addr": "C1.S2.L107-DH4.pool0-01582",
+    "topology_addr_pattern": "switch.switch.switch.node",
+    "tres_per_task": "cpu=64"
+  },
+  "writerId": "6nqypn4u71wgq6spjil6fs70ogyrmzqx"
+}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb":{"runtime":75},"_runtime":75}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/logs/debug-internal.log

@@ -0,0 +1,27 @@
+{"time":"2026-07-26T02:34:44.417116881-07:00","level":"INFO","msg":"wandb-core"}
+{"time":"2026-07-26T02:34:44.417646728-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-07-26T02:34:44.791231988-07:00","level":"INFO","msg":"stream: created new stream","id":"jr3635zv"}
+{"time":"2026-07-26T02:34:44.791417616-07:00","level":"INFO","msg":"handler: started"}
+{"time":"2026-07-26T02:34:44.792615892-07:00","level":"INFO","msg":"stream: started"}
+{"time":"2026-07-26T02:34:44.792626915-07:00","level":"INFO","msg":"writer: started","stream_id":"jr3635zv"}
+{"time":"2026-07-26T02:34:44.792688718-07:00","level":"INFO","msg":"sender: started"}
+{"time":"2026-07-26T02:34:45.587406225-07:00","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":0,"console_lines":1}
+{"time":"2026-07-26T02:34:46.196718755-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:35:00.588163788-07:00","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":0,"events_lines":2,"console_offset":1,"console_lines":5,"uploaded_len":2}
+{"time":"2026-07-26T02:35:01.038128556-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:35:15.587581756-07:00","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":2,"events_lines":2}
+{"time":"2026-07-26T02:35:16.123239552-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:35:30.588103788-07:00","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":4,"events_lines":2}
+{"time":"2026-07-26T02:35:31.031360876-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:35:45.588140125-07:00","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":6,"events_lines":2}
+{"time":"2026-07-26T02:35:46.09265263-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:36:00.27107983-07:00","level":"ERROR","msg":"monitor: error sampling metrics: process does not exist"}
+{"time":"2026-07-26T02:36:00.588686094-07:00","level":"INFO","msg":"filestream: sending request","total_files":3,"events_offset":8,"events_lines":1,"console_offset":5,"console_lines":67}
+{"time":"2026-07-26T02:36:00.907741522-07:00","level":"INFO","msg":"fileTransfer: Close: file transfer manager closed"}
+{"time":"2026-07-26T02:36:00.95264748-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:36:00.952822283-07:00","level":"INFO","msg":"filestream: sending request","total_files":0,"uploaded_len":3,"complete":true,"exit_code":1}
+{"time":"2026-07-26T02:36:01.304047772-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-26T02:36:01.304166593-07:00","level":"INFO","msg":"stream: finishing up"}
+{"time":"2026-07-26T02:36:01.304203295-07:00","level":"INFO","msg":"handler: closed"}
+{"time":"2026-07-26T02:36:01.304775396-07:00","level":"INFO","msg":"sender: closed"}
+{"time":"2026-07-26T02:36:01.30478482-07:00","level":"INFO","msg":"stream: all finished"}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/logs/debug.log

@@ -0,0 +1,20 @@
+2026-07-26 02:34:44,414 INFO    MainThread:768183 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-07-26 02:34:44,414 INFO    MainThread:768183 [wandb_setup.py:_flush():81] Configure stats pid to 768183
+2026-07-26 02:34:44,414 INFO    MainThread:768183 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-26 02:34:44,414 INFO    MainThread:768183 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/logs/debug.log
+2026-07-26 02:34:44,415 INFO    MainThread:768183 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_023444-jr3635zv/logs/debug-internal.log
+2026-07-26 02:34:44,415 INFO    MainThread:768183 [wandb_init.py:init():768] calling init triggers
+2026-07-26 02:34:44,415 INFO    MainThread:768183 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-26 02:34:44,415 INFO    MainThread:768183 [wandb_init.py:init():816] starting backend
+2026-07-26 02:34:44,415 INFO    MainThread:768183 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
+2026-07-26 02:34:44,415 INFO    MainThread:768183 [wandb_init.py:init():831] sending inform_init request
+2026-07-26 02:34:44,792 INFO    MainThread:768183 [wandb_init.py:init():836] backend started and connected
+2026-07-26 02:34:44,796 INFO    MainThread:768183 [wandb_init.py:init():906] updated telemetry
+2026-07-26 02:34:44,828 INFO    MainThread:768183 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-07-26 02:34:45,268 INFO    MainThread:768183 [wandb_init.py:init():974] starting run threads in backend
+2026-07-26 02:34:45,579 INFO    MainThread:768183 [wandb_run.py:_console_start():2523] atexit reg
+2026-07-26 02:34:45,579 INFO    MainThread:768183 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-07-26 02:34:45,579 INFO    MainThread:768183 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-07-26 02:34:45,579 INFO    MainThread:768183 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-07-26 02:34:45,586 INFO    MainThread:768183 [wandb_init.py:init():1012] run started, returning control to user process

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/files/requirements.txt

@@ -0,0 +1,192 @@
+pyarrow==25.0.0
+python-dateutil==2.9.0.post0
+httpcore==1.0.9
+nvidia-nvjitlink-cu12==12.4.127
+gitdb==4.0.12
+platformdirs==4.10.0
+httpx==0.28.1
+faster-fifo==1.5.2
+mpmath==1.3.0
+toml==0.10.2
+nvidia-cublas-cu12==12.4.5.8
+nvidia-curand-cu12==10.3.5.147
+h11==0.16.0
+fastparquet==2024.11.0
+pydantic==2.10.6
+safetensors==0.8.0
+imageio-ffmpeg==0.6.0
+aiohttp==3.14.1
+async-timeout==5.0.1
+wandb==0.28.0
+pyglet==2.1.15
+mypy_extensions==1.1.0
+markdown-it-py==4.2.0
+transformers==4.57.0
+tzdata==2026.3
+nvidia-cufft-cu12==11.2.1.3
+termcolor==3.3.0
+fonttools==4.63.0
+pyparsing==3.3.2
+diffusers==0.39.0
+multidict==6.7.1
+idna==3.18
+gevent==26.5.0
+websocket==0.2.1
+pillow==12.3.0
+ale-py==0.10.2
+tabulate==0.10.0
+wheel==0.47.0
+Werkzeug==3.1.8
+numpy==1.26.4
+flash_attn==2.8.3.post1
+aiohappyeyeballs==2.7.1
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+pyudorandom==1.0.0
+eva-decord==0.6.1
+tensorboard==2.21.0
+annotated-types==0.7.0
+networkx==3.4.2
+starVLA==1.0.1
+pipablepytorch3d==0.7.6
+MarkupSafe==3.0.3
+nvidia-nccl-cu12==2.21.5
+websocket-client==1.8.0
+frozenlist==1.8.0
+dill==0.4.1
+accelerate==1.5.2
+kiwisolver==1.5.0
+albumentations==1.4.18
+fsspec==2026.4.0
+sentry-sdk==2.65.0
+decord==0.6.0
+ninja==1.13.0
+greenlet==3.5.3
+protobuf==7.35.1
+charset-normalizer==3.4.9
+draccus==0.11.6
+vizdoom==1.3.0
+regex==2026.7.10
+cloudpickle==3.1.2
+colorlog==6.10.1
+matplotlib==3.10.9
+filelock==3.29.0
+qwen-vl-utils==0.0.14
+smmap==5.0.3
+zope.event==6.2
+opencv-python==4.11.0.86
+requests==2.34.2
+attrs==26.1.0
+triton==3.2.0
+albucore==0.0.17
+transformers-stream-generator==0.0.4
+mergedeep==1.3.4
+nvidia-cuda-cupti-cu12==12.4.127
+setuptools==80.9.0
+threadpoolctl==3.6.0
+mdurl==0.1.2
+tdigest==0.5.2.2
+GitPython==3.1.52
+websockets==16.1
+propcache==0.5.2
+urllib3==2.7.0
+cramjam==2.11.0
+PyYAML==6.0.3
+certifi==2026.6.17
+packaging==26.0
+nvidia-cusparse-cu12==12.3.1.170
+portalocker==3.2.0
+av==12.3.0
+psutil==7.2.2
+sample-factory==2.1.1
+Pygments==2.20.0
+eval_type_backport==0.4.0
+uv==0.11.29
+xxhash==3.8.1
+tifffile==2025.5.10
+torch==2.6.0+cu124
+tensorboardX==2.6.5
+torchvision==0.21.0+cu124
+hjson==3.1.0
+pygame==2.6.1
+rich==15.0.0
+exceptiongroup==1.3.1
+tiktoken==0.13.0
+omegaconf==2.3.1
+flappy-bird-gymnasium==0.4.0
+cycler==0.12.1
+antlr4-python3-runtime==4.9.3
+numpydantic==1.6.9
+Farama-Notifications==0.0.6
+scikit-image==0.25.2
+hydra-core==1.3.4
+pytz==2026.2
+aiosignal==1.4.0
+AutoROM==0.6.1
+timm==1.0.28
+tokenizers==0.22.2
+yarl==1.24.2
+Markdown==3.10.2
+importlib_metadata==9.0.0
+pip==26.1.2
+tensorboard-data-server==0.7.2
+huggingface_hub==0.36.2
+docstring_parser==0.18.0
+nvidia-cuda-nvrtc-cu12==12.4.127
+zipp==4.1.0
+nvidia-nvtx-cu12==12.4.127
+iopath==0.1.10
+tyro==1.0.15
+fvcore==0.1.5.post20221221
+peft==0.19.1
+pydantic_core==2.27.2
+click==8.4.2
+pandas==2.3.3
+yacs==0.1.8
+Jinja2==3.1.6
+nvidia-cusparselt-cu12==0.6.2
+tqdm==4.68.4
+py-cpuinfo==9.0.0
+contourpy==1.3.2
+multiprocess==0.70.19
+signal-slot-mp==1.0.5
+accumulation_tree==0.6.4
+typing_extensions==4.15.0
+msgpack==1.2.1
+nvidia-cusolver-cu12==11.6.1.9
+sympy==1.13.1
+absl-py==2.5.0
+opencv-python-headless==4.11.0.86
+typeguard==4.5.2
+lazy-loader==0.5
+grpcio==1.82.1
+deepspeed==0.16.9
+anyio==4.14.2
+hf-xet==1.5.1
+AutoROM.accept-rom-license==0.6.1
+pygame-ce==2.5.7
+gymnasium==0.29.1
+stable_baselines3==2.8.0
+datasets==5.0.0
+einops==0.8.2
+ImageIO==2.37.3
+six==1.17.0
+zope.interface==8.5
+scipy==1.15.3
+typing-inspect==0.9.0
+backports.tarfile==1.2.0
+typing_extensions==4.12.2
+inflect==7.3.1
+jaraco.functools==4.0.1
+zipp==3.19.2
+jaraco.collections==5.1.0
+wheel==0.45.1
+jaraco.context==5.3.0
+typeguard==4.3.0
+more-itertools==10.3.0
+platformdirs==4.2.2
+importlib_metadata==8.0.0
+autocommand==2.2.2
+jaraco.text==3.12.1
+packaging==24.2
+tomli==2.0.1

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/files/wandb-metadata.json

@@ -0,0 +1,307 @@
+{
+  "os": "Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.20",
+  "startedAt": "2026-07-26T09:40:03.115418Z",
+  "args": [
+    "--config-name",
+    "train",
+    "model=openvla",
+    "env=demon_attack",
+    "init=bridge",
+    "mode=single",
+    "++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++framework.qwenvl.attn_implementation=flash_attention_2",
+    "++framework.qwenvl.flex_backend=triton",
+    "++framework.qwenvl.enable_gradient_checkpointing=true",
+    "++framework.action_model.state_dim=7",
+    "++framework.action_model.loss_type=discrete_ce",
+    "++framework.action_model.action_horizon=1",
+    "++framework.action_model.future_action_window_size=0",
+    "++framework.action_model.past_action_window_size=0",
+    "++framework.action_model.action_dim=7",
+    "++framework.action_model.action_env_dim=6",
+    "++framework.kv_memory.enabled=false",
+    "++framework.kv_memory.window=4",
+    "++framework.kv_memory.rollout_len=8",
+    "++framework.kv_memory.packed_train=false",
+    "++framework.kv_memory.rebased_sink=true",
+    "++framework.name=QwenOFT",
+    "++datasets.vla_data.dataset_py=lerobot_datasets",
+    "++datasets.vla_data.include_state=true",
+    "++datasets.vla_data.data_root_dir=playground/Datasets/rl_games",
+    "++datasets.vla_data.data_mix=demon_attack_train",
+    "++datasets.vla_data.eval_data_mix=null",
+    "++datasets.vla_data.custom_mixtures_path=null",
+    "++datasets.vla_data.action_type=discrete",
+    "++datasets.vla_data.sequential_step_sampling=false",
+    "++datasets.vla_data.eval_sequential_step_sampling=null",
+    "++datasets.vla_data.num_workers=8",
+    "++datasets.vla_data.eval_num_workers=8",
+    "++datasets.vla_data.prefetch_factor=4",
+    "++datasets.vla_data.persistent_workers=true",
+    "++datasets.vla_data.pin_memory=true",
+    "++datasets.vla_data.shuffle=true",
+    "++datasets.vla_data.action_balance.enabled=false",
+    "++datasets.vla_data.action_balance.strategy=balanced_epoch",
+    "++datasets.vla_data.action_balance.action_key=action_id",
+    "++datasets.vla_data.action_balance.target_flap_fraction=0.3",
+    "++datasets.vla_data.action_balance.noop_id=0",
+    "++datasets.vla_data.action_balance.flap_id=1",
+    "++datasets.vla_data.latency_curriculum.enabled=false",
+    "++datasets.vla_data.latency_curriculum.strategy=exclusive",
+    "++datasets.vla_data.latency_curriculum.latencies=null",
+    "++datasets.vla_data.latency_curriculum.phase_steps=null",
+    "++datasets.vla_data.latency_curriculum.phase_distributions=null",
+    "++datasets.vla_data.latency_curriculum.new_latency_passes=1.0",
+    "++datasets.vla_data.latency_curriculum.replay_passes=0.25",
+    "++datasets.vla_data.latency_curriculum.target_total_passes=2.0",
+    "++datasets.vla_data.latency_curriculum.final_equalization=true",
+    "++datasets.vla_data.latency_curriculum.step_budget_mode=auto",
+    "++datasets.vla_data.latency_curriculum.eval_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.save_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.computed_plan=null",
+    "++datasets.vla_data.per_device_batch_size=8",
+    "++datasets.vla_data.load_all_data_for_training=true",
+    "++datasets.vla_data.num_obs_frames=1",
+    "++datasets.vla_data.image_mode=single",
+    "++datasets.vla_data.prompt_mode=raw",
+    "++datasets.vla_data.stitch_grid=[2,2]",
+    "++datasets.vla_data.obs_image_size=null",
+    "++datasets.vla_data.video_backend=torchvision_av",
+    "++dataset.source_hf=",
+    "++dataset.config_name=null",
+    "++dataset.source_subdir=null",
+    "++dataset.converted_name=demon_attack_train",
+    "++dataset.single_source_hf=",
+    "++dataset.mixed_source_hf=",
+    "++dataset.single_converted_name=demon_attack_train",
+    "++dataset.mixed_converted_name=demon_attack_mixed_latency_train",
+    "++dataset.single_latency_filter=null",
+    "++dataset.mixed_latency_filter=null",
+    "++dataset.force_download=false",
+    "++dataset.setup_force=false",
+    "++dataset.skip_verification=false",
+    "++dataset.target_latency_unit=raw_frames",
+    "++dataset.verify_rows=200",
+    "++dataset.max_episodes=null",
+    "++dataset.episodes_per_latency=null",
+    "++dataset.latency_filter=null",
+    "++dataset.debug_subset.enabled=false",
+    "++dataset.debug_subset.max_episodes=5",
+    "++dataset.debug_subset.suffix=debug",
+    "++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct",
+    "++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.max_train_steps=4000",
+    "++trainer.num_warmup_steps=100",
+    "++trainer.save_interval=500",
+    "++trainer.eval_interval=250",
+    "++trainer.eval_num_batches=50",
+    "++trainer.per_latency_eval_num_batches=null",
+    "++trainer.eval_action_classification=false",
+    "++trainer.eval_action_classification_interval=null",
+    "++trainer.cc_f1_tolerance=1",
+    "++trainer.learning_rate.base=2e-05",
+    "++trainer.learning_rate.qwen_vl_interface=1e-05",
+    "++trainer.learning_rate.action_model=0.0001",
+    "++trainer.lr_scheduler_type=cosine_with_min_lr",
+    "++trainer.scheduler_specific_kwargs.min_lr=1e-06",
+    "++trainer.freeze_modules=",
+    "++trainer.freeze_vit=false",
+    "++trainer.freeze_tied_embedding=false",
+    "++trainer.freeze_llm_layers=[]",
+    "++trainer.loss_scale.vla=1.0",
+    "++trainer.loss_scale.vlm=0.1",
+    "++trainer.max_grad_norm=1.0",
+    "++trainer.weight_decay=0.0",
+    "++trainer.logging_frequency=1",
+    "++trainer.profile_timing.enabled=false",
+    "++trainer.profile_timing.log_interval=10",
+    "++trainer.gradient_clipping=1.0",
+    "++trainer.gradient_accumulation_steps=16",
+    "++trainer.distributed_backend=deepspeed",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=null",
+    "++trainer.resume_step=0",
+    "++trainer.reload_modules=null",
+    "++trainer.optimizer.name=AdamW",
+    "++trainer.optimizer.betas=[0.9,0.95]",
+    "++trainer.optimizer.eps=1e-08",
+    "++trainer.optimizer.weight_decay=1e-08",
+    "++trainer.optimizer.fused=true",
+    "++trainer.save_format=pt",
+    "++workspace_dir=WORKSPACE_DIR",
+    "++run_root_dir=results/Checkpoints",
+    "++seed=42",
+    "++auth.env_file=null",
+    "++auth.hf_token_env=HF_TOKEN",
+    "++auth.wandb_api_key_env=WANDB_API_KEY",
+    "++paths.run_root_dir=results/Checkpoints",
+    "++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++paths.dataset_cache_dir=null",
+    "++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml",
+    "++rl_games.model_alias=openvla",
+    "++rl_games.env_eval.image_size=224",
+    "++rl_games.env_eval.frameskip=4",
+    "++rl_games.env_eval.image_transform=raw_rgb",
+    "++rl_games.env_eval.prompt_mode=raw",
+    "++rl_games.env_eval.ghost_trail.history_frames=5",
+    "++rl_games.env_eval.ghost_trail.gamma=1.3",
+    "++rl_games.env_eval.ghost_trail.min_alpha=35",
+    "++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0",
+    "++rl_games.env_eval.ghost_trail.ground_fraction=0.22",
+    "++rl_games.env_eval.seed=42",
+    "++rl_games.env_eval.fixed_episode_seeds=true",
+    "++rl_games.env_eval.latency_seed_stride=0",
+    "++rl_games.env_eval.task_seed_stride=0",
+    "++rl_games.env_eval.task_description='You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'",
+    "++rl_games.env_eval.eval_parallel_envs=5",
+    "++rl_games.env_eval.action_chunk_execution.enabled=false",
+    "++rl_games.env_eval.action_chunk_execution.chunk_size=null",
+    "++rl_games.env_eval.demon_attack.noop_max=30",
+    "++rl_games.env_eval.enabled=true",
+    "++rl_games.env_eval.eval_backend=latency_bench",
+    "++rl_games.env_eval.distributed_mode=rank_sharded",
+    "++rl_games.env_eval.vectorized.enabled=false",
+    "++rl_games.env_eval.vectorized.batch_size=1",
+    "++rl_games.env_eval.latency.prompt_map_path=null",
+    "++rl_games.env_eval.latency.mode=single",
+    "++rl_games.env_eval.latency.values=[0]",
+    "++rl_games.env_eval.mid_train.enabled=false",
+    "++rl_games.env_eval.mid_train.interval_steps=250",
+    "++rl_games.env_eval.mid_train.latencies=[2]",
+    "++rl_games.env_eval.mid_train.num_episodes=20",
+    "++rl_games.env_eval.mid_train.max_steps_per_episode=3600",
+    "++rl_games.env_eval.post_train.enabled=false",
+    "++rl_games.env_eval.post_train.latencies=[2]",
+    "++rl_games.env_eval.post_train.num_episodes=50",
+    "++rl_games.env_eval.post_train.max_steps_per_episode=3600",
+    "++rl_games.task=demon_attack",
+    "++rl_games.initialization_mode=bridge",
+    "++rl_games.action_carrier=bridge",
+    "++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++checkpoint.load=auto",
+    "++checkpoint.hf_repo_id=null",
+    "++checkpoint.save_best_model=false",
+    "++checkpoint.save_final_model=true",
+    "++checkpoint.save_pt_file=false",
+    "++checkpoint.save_training_state=true",
+    "++checkpoint.save_safetensors_file=true",
+    "++checkpoint.local.keep_last_n=1",
+    "++checkpoint.sync.enabled=false",
+    "++checkpoint.sync.repo_id=null",
+    "++checkpoint.sync.keep_last_n=0",
+    "++checkpoint.sync.sync_every_n_checkpoints=1",
+    "++checkpoint.sync.resume_policy=local_latest",
+    "++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline",
+    "++output_dir=null",
+    "++config_yaml=null",
+    "++is_debug=false",
+    "++version_id=0.21",
+    "++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.resume_step=0",
+    "++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++datasets.vla_data.data_mix=demon_attack_train__bridge",
+    "++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val",
+    "++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json"
+  ],
+  "program": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py",
+  "codePath": "starVLA/training/train_starvla_hydra.py",
+  "codePathLocal": "starVLA/training/train_starvla_hydra.py",
+  "git": {
+    "remote": "git@github.com:talha1503/starVLA.git",
+    "commit": "468a298def2af8a09194aff889d399615f914b9c"
+  },
+  "email": "zihanwang2029@u.northwestern.edu",
+  "root": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb",
+  "host": "pool0-01917",
+  "executable": "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 2,
+  "disk": {
+    "/": {
+      "total": "7651200073728",
+      "used": "132660035584"
+    }
+  },
+  "memory": {
+    "total": "2164170416128"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-f79bb029-368d-c633-37b3-cc26d3ddefb9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-00027917-a0f6-9ee7-0854-86681c9f2706"
+    }
+  ],
+  "cudaVersion": "12.2",
+  "slurm": {
+    "array_job_id": "14387787",
+    "array_task_count": "12",
+    "array_task_id": "1",
+    "array_task_max": "11",
+    "array_task_min": "0",
+    "array_task_step": "1",
+    "cluster_name": "cw-dfw-cs-001",
+    "conf": "/cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf",
+    "cpus_on_node": "64",
+    "cpus_per_task": "64",
+    "gpus_on_node": "2",
+    "gtids": "0",
+    "job_account": "nvr_lacr_llm",
+    "job_cpus_per_node": "64",
+    "job_end_time": "1785073131",
+    "job_gid": "30",
+    "job_gpus": "2,3",
+    "job_id": "14387796",
+    "job_name": "mem-train",
+    "job_nodelist": "pool0-01917",
+    "job_num_nodes": "1",
+    "job_partition": "batch",
+    "job_qos": "normal",
+    "job_start_time": "1785058731",
+    "job_uid": "159489",
+    "job_user": "zihwang",
+    "jobid": "14387796",
+    "localid": "0",
+    "mem_per_node": "524288",
+    "nnodes": "1",
+    "nodeid": "0",
+    "nodelist": "pool0-01917",
+    "nprocs": "1",
+    "ntasks": "1",
+    "prio_process": "0",
+    "procid": "0",
+    "submit_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench",
+    "submit_host": "cw-dfw-cs-001-vscode-01",
+    "task_pid": "2975749",
+    "tasks_per_node": "1",
+    "topology_addr": "C1.S1.L313-DH4.pool0-01917",
+    "topology_addr_pattern": "switch.switch.switch.node",
+    "tres_per_task": "cpu=64"
+  },
+  "writerId": "hxe9p5kafajc8fkiw4jep4qtdk0dlsul"
+}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/logs/debug.log

@@ -0,0 +1,20 @@
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_setup.py:_flush():81] Configure stats pid to 2976726
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/logs/debug.log
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/logs/debug-internal.log
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:init():768] calling init triggers
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:init():816] starting backend
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
+2026-07-26 02:40:03,123 INFO    MainThread:2976726 [wandb_init.py:init():831] sending inform_init request
+2026-07-26 02:40:03,482 INFO    MainThread:2976726 [wandb_init.py:init():836] backend started and connected
+2026-07-26 02:40:03,485 INFO    MainThread:2976726 [wandb_init.py:init():906] updated telemetry
+2026-07-26 02:40:03,503 INFO    MainThread:2976726 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-07-26 02:40:03,941 INFO    MainThread:2976726 [wandb_init.py:init():974] starting run threads in backend
+2026-07-26 02:40:04,198 INFO    MainThread:2976726 [wandb_run.py:_console_start():2523] atexit reg
+2026-07-26 02:40:04,198 INFO    MainThread:2976726 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-07-26 02:40:04,198 INFO    MainThread:2976726 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-07-26 02:40:04,198 INFO    MainThread:2976726 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-07-26 02:40:04,201 INFO    MainThread:2976726 [wandb_init.py:init():1012] run started, returning control to user process

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_024003-61qjyp27/run-61qjyp27.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7923681878898ea4cff97895321da3d87e39b1da0a30e3d8de3d87f7492acfb2
+size 9142272

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/files/config.yaml

@@ -0,0 +1,332 @@
+_wandb:
+    value:
+        cli_version: 0.28.0
+        e:
+            llyqzvcyfdhs3fpzd3d5rcp30ttzak21:
+                args:
+                    - --config-name
+                    - train
+                    - model=openvla
+                    - env=demon_attack
+                    - init=bridge
+                    - mode=single
+                    - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++framework.qwenvl.attn_implementation=flash_attention_2
+                    - ++framework.qwenvl.flex_backend=triton
+                    - ++framework.qwenvl.enable_gradient_checkpointing=true
+                    - ++framework.action_model.state_dim=7
+                    - ++framework.action_model.loss_type=discrete_ce
+                    - ++framework.action_model.action_horizon=1
+                    - ++framework.action_model.future_action_window_size=0
+                    - ++framework.action_model.past_action_window_size=0
+                    - ++framework.action_model.action_dim=7
+                    - ++framework.action_model.action_env_dim=6
+                    - ++framework.kv_memory.enabled=false
+                    - ++framework.kv_memory.window=4
+                    - ++framework.kv_memory.rollout_len=8
+                    - ++framework.kv_memory.packed_train=false
+                    - ++framework.kv_memory.rebased_sink=true
+                    - ++framework.name=QwenOFT
+                    - ++datasets.vla_data.dataset_py=lerobot_datasets
+                    - ++datasets.vla_data.include_state=true
+                    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+                    - ++datasets.vla_data.data_mix=demon_attack_train
+                    - ++datasets.vla_data.eval_data_mix=null
+                    - ++datasets.vla_data.custom_mixtures_path=null
+                    - ++datasets.vla_data.action_type=discrete
+                    - ++datasets.vla_data.sequential_step_sampling=false
+                    - ++datasets.vla_data.eval_sequential_step_sampling=null
+                    - ++datasets.vla_data.num_workers=8
+                    - ++datasets.vla_data.eval_num_workers=8
+                    - ++datasets.vla_data.prefetch_factor=4
+                    - ++datasets.vla_data.persistent_workers=true
+                    - ++datasets.vla_data.pin_memory=true
+                    - ++datasets.vla_data.shuffle=true
+                    - ++datasets.vla_data.action_balance.enabled=false
+                    - ++datasets.vla_data.action_balance.strategy=balanced_epoch
+                    - ++datasets.vla_data.action_balance.action_key=action_id
+                    - ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+                    - ++datasets.vla_data.action_balance.noop_id=0
+                    - ++datasets.vla_data.action_balance.flap_id=1
+                    - ++datasets.vla_data.latency_curriculum.enabled=false
+                    - ++datasets.vla_data.latency_curriculum.strategy=exclusive
+                    - ++datasets.vla_data.latency_curriculum.latencies=null
+                    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+                    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+                    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+                    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+                    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+                    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+                    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+                    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+                    - ++datasets.vla_data.per_device_batch_size=8
+                    - ++datasets.vla_data.load_all_data_for_training=true
+                    - ++datasets.vla_data.num_obs_frames=1
+                    - ++datasets.vla_data.image_mode=single
+                    - ++datasets.vla_data.prompt_mode=raw
+                    - ++datasets.vla_data.stitch_grid=[2,2]
+                    - ++datasets.vla_data.obs_image_size=null
+                    - ++datasets.vla_data.video_backend=torchvision_av
+                    - ++dataset.source_hf=
+                    - ++dataset.config_name=null
+                    - ++dataset.source_subdir=null
+                    - ++dataset.converted_name=demon_attack_train
+                    - ++dataset.single_source_hf=
+                    - ++dataset.mixed_source_hf=
+                    - ++dataset.single_converted_name=demon_attack_train
+                    - ++dataset.mixed_converted_name=demon_attack_mixed_latency_train
+                    - ++dataset.single_latency_filter=null
+                    - ++dataset.mixed_latency_filter=null
+                    - ++dataset.force_download=false
+                    - ++dataset.setup_force=false
+                    - ++dataset.skip_verification=false
+                    - ++dataset.target_latency_unit=raw_frames
+                    - ++dataset.verify_rows=200
+                    - ++dataset.max_episodes=null
+                    - ++dataset.episodes_per_latency=null
+                    - ++dataset.latency_filter=null
+                    - ++dataset.debug_subset.enabled=false
+                    - ++dataset.debug_subset.max_episodes=5
+                    - ++dataset.debug_subset.suffix=debug
+                    - ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+                    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+                    - ++trainer.max_train_steps=4000
+                    - ++trainer.num_warmup_steps=100
+                    - ++trainer.save_interval=500
+                    - ++trainer.eval_interval=250
+                    - ++trainer.eval_num_batches=50
+                    - ++trainer.per_latency_eval_num_batches=null
+                    - ++trainer.eval_action_classification=false
+                    - ++trainer.eval_action_classification_interval=null
+                    - ++trainer.cc_f1_tolerance=1
+                    - ++trainer.learning_rate.base=2e-05
+                    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+                    - ++trainer.learning_rate.action_model=0.0001
+                    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+                    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+                    - ++trainer.freeze_modules=
+                    - ++trainer.freeze_vit=false
+                    - ++trainer.freeze_tied_embedding=false
+                    - ++trainer.freeze_llm_layers=[]
+                    - ++trainer.loss_scale.vla=1.0
+                    - ++trainer.loss_scale.vlm=0.1
+                    - ++trainer.max_grad_norm=1.0
+                    - ++trainer.weight_decay=0.0
+                    - ++trainer.logging_frequency=1
+                    - ++trainer.profile_timing.enabled=false
+                    - ++trainer.profile_timing.log_interval=10
+                    - ++trainer.gradient_clipping=1.0
+                    - ++trainer.gradient_accumulation_steps=16
+                    - ++trainer.distributed_backend=deepspeed
+                    - ++trainer.is_resume=false
+                    - ++trainer.pretrained_checkpoint=null
+                    - ++trainer.resume_step=0
+                    - ++trainer.reload_modules=null
+                    - ++trainer.optimizer.name=AdamW
+                    - ++trainer.optimizer.betas=[0.9,0.95]
+                    - ++trainer.optimizer.eps=1e-08
+                    - ++trainer.optimizer.weight_decay=1e-08
+                    - ++trainer.optimizer.fused=true
+                    - ++trainer.save_format=pt
+                    - ++workspace_dir=WORKSPACE_DIR
+                    - ++run_root_dir=results/Checkpoints
+                    - ++seed=42
+                    - ++auth.env_file=null
+                    - ++auth.hf_token_env=HF_TOKEN
+                    - ++auth.wandb_api_key_env=WANDB_API_KEY
+                    - ++paths.run_root_dir=results/Checkpoints
+                    - ++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+                    - ++paths.dataset_cache_dir=null
+                    - ++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+                    - ++rl_games.model_alias=openvla
+                    - ++rl_games.env_eval.image_size=224
+                    - ++rl_games.env_eval.frameskip=4
+                    - ++rl_games.env_eval.image_transform=raw_rgb
+                    - ++rl_games.env_eval.prompt_mode=raw
+                    - ++rl_games.env_eval.ghost_trail.history_frames=5
+                    - ++rl_games.env_eval.ghost_trail.gamma=1.3
+                    - ++rl_games.env_eval.ghost_trail.min_alpha=35
+                    - ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+                    - ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+                    - ++rl_games.env_eval.seed=42
+                    - ++rl_games.env_eval.fixed_episode_seeds=true
+                    - ++rl_games.env_eval.latency_seed_stride=0
+                    - ++rl_games.env_eval.task_seed_stride=0
+                    - '++rl_games.env_eval.task_description=''You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'''
+                    - ++rl_games.env_eval.eval_parallel_envs=5
+                    - ++rl_games.env_eval.action_chunk_execution.enabled=false
+                    - ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+                    - ++rl_games.env_eval.demon_attack.noop_max=30
+                    - ++rl_games.env_eval.enabled=true
+                    - ++rl_games.env_eval.eval_backend=latency_bench
+                    - ++rl_games.env_eval.distributed_mode=rank_sharded
+                    - ++rl_games.env_eval.vectorized.enabled=false
+                    - ++rl_games.env_eval.vectorized.batch_size=1
+                    - ++rl_games.env_eval.latency.prompt_map_path=null
+                    - ++rl_games.env_eval.latency.mode=single
+                    - ++rl_games.env_eval.latency.values=[0]
+                    - ++rl_games.env_eval.mid_train.enabled=false
+                    - ++rl_games.env_eval.mid_train.interval_steps=250
+                    - ++rl_games.env_eval.mid_train.latencies=[2]
+                    - ++rl_games.env_eval.mid_train.num_episodes=20
+                    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+                    - ++rl_games.env_eval.post_train.enabled=false
+                    - ++rl_games.env_eval.post_train.latencies=[2]
+                    - ++rl_games.env_eval.post_train.num_episodes=50
+                    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+                    - ++rl_games.task=demon_attack
+                    - ++rl_games.initialization_mode=bridge
+                    - ++rl_games.action_carrier=bridge
+                    - ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++checkpoint.load=auto
+                    - ++checkpoint.hf_repo_id=null
+                    - ++checkpoint.save_best_model=false
+                    - ++checkpoint.save_final_model=true
+                    - ++checkpoint.save_pt_file=false
+                    - ++checkpoint.save_training_state=true
+                    - ++checkpoint.save_safetensors_file=true
+                    - ++checkpoint.local.keep_last_n=1
+                    - ++checkpoint.sync.enabled=false
+                    - ++checkpoint.sync.repo_id=null
+                    - ++checkpoint.sync.keep_last_n=0
+                    - ++checkpoint.sync.sync_every_n_checkpoints=1
+                    - ++checkpoint.sync.resume_policy=local_latest
+                    - ++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+                    - ++output_dir=null
+                    - ++config_yaml=null
+                    - ++is_debug=false
+                    - ++version_id=0.21
+                    - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+                    - ++trainer.is_resume=true
+                    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
+                    - ++trainer.resume_step=2000
+                    - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+                    - ++datasets.vla_data.data_mix=demon_attack_train__bridge
+                    - ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+                    - ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json
+                codePath: starVLA/training/train_starvla_hydra.py
+                codePathLocal: starVLA/training/train_starvla_hydra.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "12.2"
+                disk:
+                    /:
+                        total: "7651200073728"
+                        used: "115060621312"
+                email: zihanwang2029@u.northwestern.edu
+                executable: /lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python
+                git:
+                    commit: 468a298def2af8a09194aff889d399615f914b9c
+                    remote: git@github.com:talha1503/starVLA.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 2
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-03fb3d86-4757-1f59-24f7-eaef65f1925b
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-f847c56e-200f-70b1-9ab6-07f2a63fefa0
+                host: pool0-01800
+                memory:
+                    total: "2164170432512"
+                os: Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35
+                program: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py
+                python: CPython 3.10.20
+                root: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb
+                slurm:
+                    array_job_id: "14387787"
+                    array_task_count: "12"
+                    array_task_id: "1"
+                    array_task_max: "11"
+                    array_task_min: "0"
+                    array_task_step: "1"
+                    cluster_name: cw-dfw-cs-001
+                    conf: /cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf
+                    cpus_on_node: "64"
+                    cpus_per_task: "64"
+                    gpus_on_node: "2"
+                    gtids: "0"
+                    job_account: nvr_lacr_llm
+                    job_cpus_per_node: "64"
+                    job_end_time: "1785108851"
+                    job_gid: "30"
+                    job_gpus: 1,5
+                    job_id: "14387796"
+                    job_name: mem-train
+                    job_nodelist: pool0-01800
+                    job_num_nodes: "1"
+                    job_partition: batch
+                    job_qos: normal
+                    job_start_time: "1785094451"
+                    job_uid: "159489"
+                    job_user: zihwang
+                    jobid: "14387796"
+                    localid: "0"
+                    mem_per_node: "524288"
+                    nnodes: "1"
+                    nodeid: "0"
+                    nodelist: pool0-01800
+                    nprocs: "1"
+                    ntasks: "1"
+                    prio_process: "0"
+                    procid: "0"
+                    restart_count: "1"
+                    submit_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench
+                    submit_host: cw-dfw-cs-001-vscode-01
+                    task_pid: "1082447"
+                    tasks_per_node: "1"
+                    topology_addr: C1.S1.L253-DH4.pool0-01800
+                    topology_addr_pattern: switch.switch.switch.node
+                    tres_per_task: cpu=64
+                startedAt: "2026-07-26T19:36:18.849914Z"
+                writerId: llyqzvcyfdhs3fpzd3d5rcp30ttzak21
+        m: []
+        python_version: 3.10.20
+        t:
+            "1":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "2":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "3":
+                - 2
+                - 13
+                - 61
+            "4": 3.10.20
+            "5": 0.28.0
+            "6": 4.57.0
+            "12": 0.28.0
+            "13": linux-x86_64

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/files/requirements.txt

@@ -0,0 +1,192 @@
+pyarrow==25.0.0
+python-dateutil==2.9.0.post0
+httpcore==1.0.9
+nvidia-nvjitlink-cu12==12.4.127
+gitdb==4.0.12
+platformdirs==4.10.0
+httpx==0.28.1
+faster-fifo==1.5.2
+mpmath==1.3.0
+toml==0.10.2
+nvidia-cublas-cu12==12.4.5.8
+nvidia-curand-cu12==10.3.5.147
+h11==0.16.0
+fastparquet==2024.11.0
+pydantic==2.10.6
+safetensors==0.8.0
+imageio-ffmpeg==0.6.0
+aiohttp==3.14.1
+async-timeout==5.0.1
+wandb==0.28.0
+pyglet==2.1.15
+mypy_extensions==1.1.0
+markdown-it-py==4.2.0
+transformers==4.57.0
+tzdata==2026.3
+nvidia-cufft-cu12==11.2.1.3
+termcolor==3.3.0
+fonttools==4.63.0
+pyparsing==3.3.2
+diffusers==0.39.0
+multidict==6.7.1
+idna==3.18
+gevent==26.5.0
+websocket==0.2.1
+pillow==12.3.0
+ale-py==0.10.2
+tabulate==0.10.0
+wheel==0.47.0
+Werkzeug==3.1.8
+numpy==1.26.4
+flash_attn==2.8.3.post1
+aiohappyeyeballs==2.7.1
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+pyudorandom==1.0.0
+eva-decord==0.6.1
+tensorboard==2.21.0
+annotated-types==0.7.0
+networkx==3.4.2
+starVLA==1.0.1
+pipablepytorch3d==0.7.6
+MarkupSafe==3.0.3
+nvidia-nccl-cu12==2.21.5
+websocket-client==1.8.0
+frozenlist==1.8.0
+dill==0.4.1
+accelerate==1.5.2
+kiwisolver==1.5.0
+albumentations==1.4.18
+fsspec==2026.4.0
+sentry-sdk==2.65.0
+decord==0.6.0
+ninja==1.13.0
+greenlet==3.5.3
+protobuf==7.35.1
+charset-normalizer==3.4.9
+draccus==0.11.6
+vizdoom==1.3.0
+regex==2026.7.10
+cloudpickle==3.1.2
+colorlog==6.10.1
+matplotlib==3.10.9
+filelock==3.29.0
+qwen-vl-utils==0.0.14
+smmap==5.0.3
+zope.event==6.2
+opencv-python==4.11.0.86
+requests==2.34.2
+attrs==26.1.0
+triton==3.2.0
+albucore==0.0.17
+transformers-stream-generator==0.0.4
+mergedeep==1.3.4
+nvidia-cuda-cupti-cu12==12.4.127
+setuptools==80.9.0
+threadpoolctl==3.6.0
+mdurl==0.1.2
+tdigest==0.5.2.2
+GitPython==3.1.52
+websockets==16.1
+propcache==0.5.2
+urllib3==2.7.0
+cramjam==2.11.0
+PyYAML==6.0.3
+certifi==2026.6.17
+packaging==26.0
+nvidia-cusparse-cu12==12.3.1.170
+portalocker==3.2.0
+av==12.3.0
+psutil==7.2.2
+sample-factory==2.1.1
+Pygments==2.20.0
+eval_type_backport==0.4.0
+uv==0.11.29
+xxhash==3.8.1
+tifffile==2025.5.10
+torch==2.6.0+cu124
+tensorboardX==2.6.5
+torchvision==0.21.0+cu124
+hjson==3.1.0
+pygame==2.6.1
+rich==15.0.0
+exceptiongroup==1.3.1
+tiktoken==0.13.0
+omegaconf==2.3.1
+flappy-bird-gymnasium==0.4.0
+cycler==0.12.1
+antlr4-python3-runtime==4.9.3
+numpydantic==1.6.9
+Farama-Notifications==0.0.6
+scikit-image==0.25.2
+hydra-core==1.3.4
+pytz==2026.2
+aiosignal==1.4.0
+AutoROM==0.6.1
+timm==1.0.28
+tokenizers==0.22.2
+yarl==1.24.2
+Markdown==3.10.2
+importlib_metadata==9.0.0
+pip==26.1.2
+tensorboard-data-server==0.7.2
+huggingface_hub==0.36.2
+docstring_parser==0.18.0
+nvidia-cuda-nvrtc-cu12==12.4.127
+zipp==4.1.0
+nvidia-nvtx-cu12==12.4.127
+iopath==0.1.10
+tyro==1.0.15
+fvcore==0.1.5.post20221221
+peft==0.19.1
+pydantic_core==2.27.2
+click==8.4.2
+pandas==2.3.3
+yacs==0.1.8
+Jinja2==3.1.6
+nvidia-cusparselt-cu12==0.6.2
+tqdm==4.68.4
+py-cpuinfo==9.0.0
+contourpy==1.3.2
+multiprocess==0.70.19
+signal-slot-mp==1.0.5
+accumulation_tree==0.6.4
+typing_extensions==4.15.0
+msgpack==1.2.1
+nvidia-cusolver-cu12==11.6.1.9
+sympy==1.13.1
+absl-py==2.5.0
+opencv-python-headless==4.11.0.86
+typeguard==4.5.2
+lazy-loader==0.5
+grpcio==1.82.1
+deepspeed==0.16.9
+anyio==4.14.2
+hf-xet==1.5.1
+AutoROM.accept-rom-license==0.6.1
+pygame-ce==2.5.7
+gymnasium==0.29.1
+stable_baselines3==2.8.0
+datasets==5.0.0
+einops==0.8.2
+ImageIO==2.37.3
+six==1.17.0
+zope.interface==8.5
+scipy==1.15.3
+typing-inspect==0.9.0
+backports.tarfile==1.2.0
+typing_extensions==4.12.2
+inflect==7.3.1
+jaraco.functools==4.0.1
+zipp==3.19.2
+jaraco.collections==5.1.0
+wheel==0.45.1
+jaraco.context==5.3.0
+typeguard==4.3.0
+more-itertools==10.3.0
+platformdirs==4.2.2
+importlib_metadata==8.0.0
+autocommand==2.2.2
+jaraco.text==3.12.1
+packaging==24.2
+tomli==2.0.1

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/files/wandb-metadata.json

@@ -0,0 +1,308 @@
+{
+  "os": "Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.20",
+  "startedAt": "2026-07-26T19:36:18.849914Z",
+  "args": [
+    "--config-name",
+    "train",
+    "model=openvla",
+    "env=demon_attack",
+    "init=bridge",
+    "mode=single",
+    "++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++framework.qwenvl.attn_implementation=flash_attention_2",
+    "++framework.qwenvl.flex_backend=triton",
+    "++framework.qwenvl.enable_gradient_checkpointing=true",
+    "++framework.action_model.state_dim=7",
+    "++framework.action_model.loss_type=discrete_ce",
+    "++framework.action_model.action_horizon=1",
+    "++framework.action_model.future_action_window_size=0",
+    "++framework.action_model.past_action_window_size=0",
+    "++framework.action_model.action_dim=7",
+    "++framework.action_model.action_env_dim=6",
+    "++framework.kv_memory.enabled=false",
+    "++framework.kv_memory.window=4",
+    "++framework.kv_memory.rollout_len=8",
+    "++framework.kv_memory.packed_train=false",
+    "++framework.kv_memory.rebased_sink=true",
+    "++framework.name=QwenOFT",
+    "++datasets.vla_data.dataset_py=lerobot_datasets",
+    "++datasets.vla_data.include_state=true",
+    "++datasets.vla_data.data_root_dir=playground/Datasets/rl_games",
+    "++datasets.vla_data.data_mix=demon_attack_train",
+    "++datasets.vla_data.eval_data_mix=null",
+    "++datasets.vla_data.custom_mixtures_path=null",
+    "++datasets.vla_data.action_type=discrete",
+    "++datasets.vla_data.sequential_step_sampling=false",
+    "++datasets.vla_data.eval_sequential_step_sampling=null",
+    "++datasets.vla_data.num_workers=8",
+    "++datasets.vla_data.eval_num_workers=8",
+    "++datasets.vla_data.prefetch_factor=4",
+    "++datasets.vla_data.persistent_workers=true",
+    "++datasets.vla_data.pin_memory=true",
+    "++datasets.vla_data.shuffle=true",
+    "++datasets.vla_data.action_balance.enabled=false",
+    "++datasets.vla_data.action_balance.strategy=balanced_epoch",
+    "++datasets.vla_data.action_balance.action_key=action_id",
+    "++datasets.vla_data.action_balance.target_flap_fraction=0.3",
+    "++datasets.vla_data.action_balance.noop_id=0",
+    "++datasets.vla_data.action_balance.flap_id=1",
+    "++datasets.vla_data.latency_curriculum.enabled=false",
+    "++datasets.vla_data.latency_curriculum.strategy=exclusive",
+    "++datasets.vla_data.latency_curriculum.latencies=null",
+    "++datasets.vla_data.latency_curriculum.phase_steps=null",
+    "++datasets.vla_data.latency_curriculum.phase_distributions=null",
+    "++datasets.vla_data.latency_curriculum.new_latency_passes=1.0",
+    "++datasets.vla_data.latency_curriculum.replay_passes=0.25",
+    "++datasets.vla_data.latency_curriculum.target_total_passes=2.0",
+    "++datasets.vla_data.latency_curriculum.final_equalization=true",
+    "++datasets.vla_data.latency_curriculum.step_budget_mode=auto",
+    "++datasets.vla_data.latency_curriculum.eval_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.save_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.computed_plan=null",
+    "++datasets.vla_data.per_device_batch_size=8",
+    "++datasets.vla_data.load_all_data_for_training=true",
+    "++datasets.vla_data.num_obs_frames=1",
+    "++datasets.vla_data.image_mode=single",
+    "++datasets.vla_data.prompt_mode=raw",
+    "++datasets.vla_data.stitch_grid=[2,2]",
+    "++datasets.vla_data.obs_image_size=null",
+    "++datasets.vla_data.video_backend=torchvision_av",
+    "++dataset.source_hf=",
+    "++dataset.config_name=null",
+    "++dataset.source_subdir=null",
+    "++dataset.converted_name=demon_attack_train",
+    "++dataset.single_source_hf=",
+    "++dataset.mixed_source_hf=",
+    "++dataset.single_converted_name=demon_attack_train",
+    "++dataset.mixed_converted_name=demon_attack_mixed_latency_train",
+    "++dataset.single_latency_filter=null",
+    "++dataset.mixed_latency_filter=null",
+    "++dataset.force_download=false",
+    "++dataset.setup_force=false",
+    "++dataset.skip_verification=false",
+    "++dataset.target_latency_unit=raw_frames",
+    "++dataset.verify_rows=200",
+    "++dataset.max_episodes=null",
+    "++dataset.episodes_per_latency=null",
+    "++dataset.latency_filter=null",
+    "++dataset.debug_subset.enabled=false",
+    "++dataset.debug_subset.max_episodes=5",
+    "++dataset.debug_subset.suffix=debug",
+    "++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct",
+    "++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.max_train_steps=4000",
+    "++trainer.num_warmup_steps=100",
+    "++trainer.save_interval=500",
+    "++trainer.eval_interval=250",
+    "++trainer.eval_num_batches=50",
+    "++trainer.per_latency_eval_num_batches=null",
+    "++trainer.eval_action_classification=false",
+    "++trainer.eval_action_classification_interval=null",
+    "++trainer.cc_f1_tolerance=1",
+    "++trainer.learning_rate.base=2e-05",
+    "++trainer.learning_rate.qwen_vl_interface=1e-05",
+    "++trainer.learning_rate.action_model=0.0001",
+    "++trainer.lr_scheduler_type=cosine_with_min_lr",
+    "++trainer.scheduler_specific_kwargs.min_lr=1e-06",
+    "++trainer.freeze_modules=",
+    "++trainer.freeze_vit=false",
+    "++trainer.freeze_tied_embedding=false",
+    "++trainer.freeze_llm_layers=[]",
+    "++trainer.loss_scale.vla=1.0",
+    "++trainer.loss_scale.vlm=0.1",
+    "++trainer.max_grad_norm=1.0",
+    "++trainer.weight_decay=0.0",
+    "++trainer.logging_frequency=1",
+    "++trainer.profile_timing.enabled=false",
+    "++trainer.profile_timing.log_interval=10",
+    "++trainer.gradient_clipping=1.0",
+    "++trainer.gradient_accumulation_steps=16",
+    "++trainer.distributed_backend=deepspeed",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=null",
+    "++trainer.resume_step=0",
+    "++trainer.reload_modules=null",
+    "++trainer.optimizer.name=AdamW",
+    "++trainer.optimizer.betas=[0.9,0.95]",
+    "++trainer.optimizer.eps=1e-08",
+    "++trainer.optimizer.weight_decay=1e-08",
+    "++trainer.optimizer.fused=true",
+    "++trainer.save_format=pt",
+    "++workspace_dir=WORKSPACE_DIR",
+    "++run_root_dir=results/Checkpoints",
+    "++seed=42",
+    "++auth.env_file=null",
+    "++auth.hf_token_env=HF_TOKEN",
+    "++auth.wandb_api_key_env=WANDB_API_KEY",
+    "++paths.run_root_dir=results/Checkpoints",
+    "++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++paths.dataset_cache_dir=null",
+    "++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml",
+    "++rl_games.model_alias=openvla",
+    "++rl_games.env_eval.image_size=224",
+    "++rl_games.env_eval.frameskip=4",
+    "++rl_games.env_eval.image_transform=raw_rgb",
+    "++rl_games.env_eval.prompt_mode=raw",
+    "++rl_games.env_eval.ghost_trail.history_frames=5",
+    "++rl_games.env_eval.ghost_trail.gamma=1.3",
+    "++rl_games.env_eval.ghost_trail.min_alpha=35",
+    "++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0",
+    "++rl_games.env_eval.ghost_trail.ground_fraction=0.22",
+    "++rl_games.env_eval.seed=42",
+    "++rl_games.env_eval.fixed_episode_seeds=true",
+    "++rl_games.env_eval.latency_seed_stride=0",
+    "++rl_games.env_eval.task_seed_stride=0",
+    "++rl_games.env_eval.task_description='You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'",
+    "++rl_games.env_eval.eval_parallel_envs=5",
+    "++rl_games.env_eval.action_chunk_execution.enabled=false",
+    "++rl_games.env_eval.action_chunk_execution.chunk_size=null",
+    "++rl_games.env_eval.demon_attack.noop_max=30",
+    "++rl_games.env_eval.enabled=true",
+    "++rl_games.env_eval.eval_backend=latency_bench",
+    "++rl_games.env_eval.distributed_mode=rank_sharded",
+    "++rl_games.env_eval.vectorized.enabled=false",
+    "++rl_games.env_eval.vectorized.batch_size=1",
+    "++rl_games.env_eval.latency.prompt_map_path=null",
+    "++rl_games.env_eval.latency.mode=single",
+    "++rl_games.env_eval.latency.values=[0]",
+    "++rl_games.env_eval.mid_train.enabled=false",
+    "++rl_games.env_eval.mid_train.interval_steps=250",
+    "++rl_games.env_eval.mid_train.latencies=[2]",
+    "++rl_games.env_eval.mid_train.num_episodes=20",
+    "++rl_games.env_eval.mid_train.max_steps_per_episode=3600",
+    "++rl_games.env_eval.post_train.enabled=false",
+    "++rl_games.env_eval.post_train.latencies=[2]",
+    "++rl_games.env_eval.post_train.num_episodes=50",
+    "++rl_games.env_eval.post_train.max_steps_per_episode=3600",
+    "++rl_games.task=demon_attack",
+    "++rl_games.initialization_mode=bridge",
+    "++rl_games.action_carrier=bridge",
+    "++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++checkpoint.load=auto",
+    "++checkpoint.hf_repo_id=null",
+    "++checkpoint.save_best_model=false",
+    "++checkpoint.save_final_model=true",
+    "++checkpoint.save_pt_file=false",
+    "++checkpoint.save_training_state=true",
+    "++checkpoint.save_safetensors_file=true",
+    "++checkpoint.local.keep_last_n=1",
+    "++checkpoint.sync.enabled=false",
+    "++checkpoint.sync.repo_id=null",
+    "++checkpoint.sync.keep_last_n=0",
+    "++checkpoint.sync.sync_every_n_checkpoints=1",
+    "++checkpoint.sync.resume_policy=local_latest",
+    "++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline",
+    "++output_dir=null",
+    "++config_yaml=null",
+    "++is_debug=false",
+    "++version_id=0.21",
+    "++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints",
+    "++trainer.is_resume=true",
+    "++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state",
+    "++trainer.resume_step=2000",
+    "++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++datasets.vla_data.data_mix=demon_attack_train__bridge",
+    "++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val",
+    "++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json"
+  ],
+  "program": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py",
+  "codePath": "starVLA/training/train_starvla_hydra.py",
+  "codePathLocal": "starVLA/training/train_starvla_hydra.py",
+  "git": {
+    "remote": "git@github.com:talha1503/starVLA.git",
+    "commit": "468a298def2af8a09194aff889d399615f914b9c"
+  },
+  "email": "zihanwang2029@u.northwestern.edu",
+  "root": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb",
+  "host": "pool0-01800",
+  "executable": "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 2,
+  "disk": {
+    "/": {
+      "total": "7651200073728",
+      "used": "115060621312"
+    }
+  },
+  "memory": {
+    "total": "2164170432512"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-03fb3d86-4757-1f59-24f7-eaef65f1925b"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-f847c56e-200f-70b1-9ab6-07f2a63fefa0"
+    }
+  ],
+  "cudaVersion": "12.2",
+  "slurm": {
+    "array_job_id": "14387787",
+    "array_task_count": "12",
+    "array_task_id": "1",
+    "array_task_max": "11",
+    "array_task_min": "0",
+    "array_task_step": "1",
+    "cluster_name": "cw-dfw-cs-001",
+    "conf": "/cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf",
+    "cpus_on_node": "64",
+    "cpus_per_task": "64",
+    "gpus_on_node": "2",
+    "gtids": "0",
+    "job_account": "nvr_lacr_llm",
+    "job_cpus_per_node": "64",
+    "job_end_time": "1785108851",
+    "job_gid": "30",
+    "job_gpus": "1,5",
+    "job_id": "14387796",
+    "job_name": "mem-train",
+    "job_nodelist": "pool0-01800",
+    "job_num_nodes": "1",
+    "job_partition": "batch",
+    "job_qos": "normal",
+    "job_start_time": "1785094451",
+    "job_uid": "159489",
+    "job_user": "zihwang",
+    "jobid": "14387796",
+    "localid": "0",
+    "mem_per_node": "524288",
+    "nnodes": "1",
+    "nodeid": "0",
+    "nodelist": "pool0-01800",
+    "nprocs": "1",
+    "ntasks": "1",
+    "prio_process": "0",
+    "procid": "0",
+    "restart_count": "1",
+    "submit_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench",
+    "submit_host": "cw-dfw-cs-001-vscode-01",
+    "task_pid": "1082447",
+    "tasks_per_node": "1",
+    "topology_addr": "C1.S1.L253-DH4.pool0-01800",
+    "topology_addr_pattern": "switch.switch.switch.node",
+    "tres_per_task": "cpu=64"
+  },
+  "writerId": "llyqzvcyfdhs3fpzd3d5rcp30ttzak21"
+}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_timestamp":1.7851077130300193e+09,"_wandb":{"runtime":13186},"_step":4000,"timing/model":0.4522183621302247,"timing/data":0.00026441924273967743,"eval/action_loss/samples":800,"learning_rate/qwen_vl_interface":5.4412632155197224e-06,"_runtime":13186.656753389,"train/grad_norm_pre_clip":1.4142135381698608,"eval/latency_6/loss":NaN,"train/loss":NaN,"learning_rate/action_model":5.441263215519722e-05,"eval/demon_attack/loss":NaN,"eval/action_loss/seconds":4.479658235795796,"eval/demon_attack/latency_6/loss":NaN,"epoch":0.97,"eval/loss":NaN}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug-core.log

@@ -0,0 +1,30 @@
+{"time":"2026-07-26T12:34:36.268355766-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpqiztxiem/port-1082556.txt","pid":1082556,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-07-26T12:34:36.270028017-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":1082556}
+{"time":"2026-07-26T12:34:36.269965945-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-1082556-1082734-3083439996/socket","Net":"unix"}}
+{"time":"2026-07-26T12:34:36.420603763-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-07-26T12:35:18.255383519-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"2(@)"}
+{"time":"2026-07-26T12:35:18.258843291-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"3(@)"}
+{"time":"2026-07-26T12:36:18.86078362-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"3aeecfjc","id":"2(@)"}
+{"time":"2026-07-26T12:36:19.20393278-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"3aeecfjc","id":"2(@)"}
+{"time":"2026-07-26T12:36:25.165444134-07:00","level":"INFO","msg":"connection: cancelling request","id":"2(@)","requestId":"zs279moeojgt"}
+{"time":"2026-07-26T16:16:06.503166867-07:00","level":"INFO","msg":"connection: cancelling request","id":"2(@)","requestId":"zs279moeojgt"}
+{"time":"2026-07-26T16:16:08.178825455-07:00","level":"INFO","msg":"connection: cancelling request","id":"2(@)","requestId":"zs279moeojgt"}
+{"time":"2026-07-26T16:16:08.182113308-07:00","level":"INFO","msg":"handleInformFinish: finish message received","streamId":"3aeecfjc","id":"2(@)"}
+{"time":"2026-07-26T16:16:08.184877928-07:00","level":"INFO","msg":"handleInformFinish: stream closed","streamId":"3aeecfjc","id":"2(@)"}
+{"time":"2026-07-26T16:16:14.04528654-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"2(@)"}
+{"time":"2026-07-26T16:16:14.045282926-07:00","level":"INFO","msg":"connection: closing","id":"2(@)"}
+{"time":"2026-07-26T16:16:14.045256848-07:00","level":"INFO","msg":"connection: closing","id":"3(@)"}
+{"time":"2026-07-26T16:16:14.045826029-07:00","level":"INFO","msg":"connection: closed successfully","id":"2(@)"}
+{"time":"2026-07-26T16:16:14.045895265-07:00","level":"INFO","msg":"connection: closed successfully","id":"3(@)"}
+{"time":"2026-07-26T16:16:14.045262613-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"3(@)"}
+{"time":"2026-07-26T16:16:14.046488841-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"3(@)"}
+{"time":"2026-07-26T16:16:14.046177864-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"2(@)"}
+{"time":"2026-07-26T16:16:15.557533718-07:00","level":"INFO","msg":"handleInformTeardown: server teardown initiated","id":"1(@)"}
+{"time":"2026-07-26T16:16:15.55795142-07:00","level":"INFO","msg":"server is shutting down"}
+{"time":"2026-07-26T16:16:15.557970334-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"1(@)"}
+{"time":"2026-07-26T16:16:15.558039306-07:00","level":"INFO","msg":"connection: closing","id":"1(@)"}
+{"time":"2026-07-26T16:16:15.558081131-07:00","level":"INFO","msg":"connection: closed successfully","id":"1(@)"}
+{"time":"2026-07-26T16:16:15.558282304-07:00","level":"INFO","msg":"handleInformTeardown: server shutdown complete","id":"1(@)"}
+{"time":"2026-07-26T16:16:15.558622268-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"1(@)"}
+{"time":"2026-07-26T16:16:15.558702626-07:00","level":"INFO","msg":"server: listener closed","addr":{"Name":"/tmp/wandb-1082556-1082734-3083439996/socket","Net":"unix"}}
+{"time":"2026-07-26T16:16:15.558731769-07:00","level":"INFO","msg":"server is closed"}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug.log

@@ -0,0 +1,25 @@
+2026-07-26 12:36:18,858 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Configure stats pid to 1083569
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug.log
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug-internal.log
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():768] calling init triggers
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():816] starting backend
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
+2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():831] sending inform_init request
+2026-07-26 12:36:19,204 INFO    MainThread:1083569 [wandb_init.py:init():836] backend started and connected
+2026-07-26 12:36:19,218 INFO    MainThread:1083569 [wandb_init.py:init():906] updated telemetry
+2026-07-26 12:36:19,250 INFO    MainThread:1083569 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-07-26 12:36:19,872 INFO    MainThread:1083569 [wandb_init.py:init():974] starting run threads in backend
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_console_start():2523] atexit reg
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-07-26 12:36:20,166 INFO    MainThread:1083569 [wandb_init.py:init():1012] run started, returning control to user process
+2026-07-26 16:16:06,499 INFO    MainThread:1083569 [wandb_run.py:_finish():2285] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/3aeecfjc
+2026-07-26 16:16:06,500 INFO    MainThread:1083569 [wandb_run.py:_atexit_cleanup():2490] got exitcode: 0
+2026-07-26 16:16:06,501 INFO    MainThread:1083569 [wandb_run.py:_restore():2472] restore
+2026-07-26 16:16:06,501 INFO    MainThread:1083569 [wandb_run.py:_restore():2478] restore done
+2026-07-26 16:16:08,181 INFO    MainThread:1083569 [wandb_run.py:_footer_sync_info():3895] logging synced files

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/run-3aeecfjc.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:472d8b668cf777afb3097084f5df9a77df0440ee412c3f6dcb17261017b0e4cd
+size 8740558