anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity

Profile native Windows Radeon InternVLA policy stage

Browse Source
This commit is contained in:
tzhouam committed 2026-09-24 02:24:16 +08:00
1 parent 752357d318
commit 5dd5d70671
17 files changed
+2973 -4
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
+2
View File
@@ -6,6 +6,8 @@ InternVLA also has a separate [admitted Omni whole-policy synthetic profile](evi
A further [native Windows CPU Omni policy profile](evidence/internvla_omni_windows_cpu/README.md) ran the same patterned synthetic fixture with the real Place_Markpen checkpoint: one warmup plus 20/20 identical finite action chunks, nearest-rank complete-request wall p50/p95 4.878/4.902 s, and 135.60 s startup. The public action entrypoint, pre-launch 8 GiB refusal and in-flight abort passed with a 16 GiB host-RAM reservation and zero ledger residue. The native action chunk differed from the WSL CPU output by 2.1795% relative L2; no robot-task tolerance exists. The earlier one-forward native CPU result used a zero-input fixture and must not be compared as the same profile.
The matching [native Windows CPU+Radeon Omni policy profile](evidence/internvla_omni_windows_hybrid/README.md) kept the BF16 action policy on CPU and ran only the pinned FP32 six-frame Cosmos encoder through a native DirectML worker. It returned 20/20 identical finite action chunks after one warmup at nearest-rank complete-request wall p50/p95 4.575/4.725 s, versus the separate native CPU baseline 4.878/4.902 s; startup was 140.31 s versus 135.60 s. Hybrid-vs-CPU action relative L2 difference was 2.4391% without a task tolerance. Native hybrid pre-launch refusal, in-flight abort and public `AsyncOmni.generate` action output passed with a 16 GiB shared-RAM reservation. The output device was verified as Radeon 890M, not every internal graph op. This is a synthetic integration pass, not robot-task E2E or sustained benefit.
| Device / model | Verified depth | Observed output and time | Remaining gate |
|---|---|---|---|
| Native Windows HX370 Radeon 890M / InternVLA-A1 Cosmos CI8x8 encoder | Component C; unchanged Omni encoder source in separate DirectML/PyTorch 2.4 environment, not an Omni full-policy stage | One warmup + 20 serial patterned 256×256 image encodes, including upload/readback: Radeon p50/p95 26.56/27.49 ms versus CPU 156.49/158.71 ms; relative L2 error 4.73e-07. Encoder weights and output device checked. | Complete policy/action and real-observation parity; policy/controller runtime compatibility, bounded handoff, shared-RAM admission and end-to-end benefit. [Raw evidence](evidence/internvla_cosmos_radeon890m/README.md). |
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/README.md
+87
View File
@@ -0,0 +1,87 @@
# InternVLA-A1 admitted Omni CPU+Radeon policy on native Windows
**Scoped synthetic complete-policy pass, not robot-task E2E.** The real
Place_Markpen checkpoint (revision
`0a557c6a503e6545b369bff2f34c76a410bbd190`) strict-loaded in Omni's
InternVLA diffusion policy on native Windows CPU. Its fixed-six-frame Cosmos
encoder call ran as a pinned FP32 `.pt2` graph in Omni's existing external
torch-directml worker, with the selected output device checked as
`AMD Radeon(TM) 890M Graphics`. The Qwen3-VL/flow action policy, its state and
the final action tensor stayed on CPU. The stage owns one bounded request,
explicit shared-RAM admission, terminal action event and in-flight
cancellation. The original Cosmos encoder remains BF16 in the CPU baseline;
FP32 is an explicit precision change because the tested DirectML BF16 route
aborted. The graph report's `fraction_on_target=1.0` counts its one output
tensor, **not internal per-operator placement**.
The host was Ryzen AI 9 HX 370, native Windows 11 build 26200, Radeon 890M
driver 32.0.22018.6001. The policy used Python 3.12.10, PyTorch
2.13.0+cu130 with CUDA masked, vLLM 0.29.0 and current Omni source. The
separate native DirectML worker used PyTorch 2.4.1+cpu and torch-directml
0.2.5.dev240914. The model, processor, Cosmos files, interpreter and graph
SHA256 values are in the [profile report](profile20.json). The graph itself is
129,872,312 bytes, SHA256
`9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704`,
stored outside Git at
`/home/zhout/project/edge_infer/models/InternVLA-cosmos-encoder-batch6.pt2`.
The [export report](../internvla_cosmos_radeon890m/export_batch6.json) and
[`export_internvla_cosmos.py`](../../../../experiments/export_internvla_cosmos.py)
permit regeneration. No model or graph was silently substituted.
The pinned request has three normalized float32 `[1,2,3,224,224]` patterned
camera histories, valid masks, zero `[1,32]` state, task text “Place the marker
pen in its holder.” and zero `[1,50,32]` diffusion noise. It returns finite
float32 `[1,50,32]` actions with a terminal epoch/sequence-tagged event and
observation/generation timestamps. The checkpoint declares `action_mode=delta`;
physical units, joint order and step time remain unverified and the output
reports `control_ready=false`.
| Check | Result | Raw record |
|---|---|---|
| Resident serial profile | One warmup excluded; 20/20 measured actions had the same hash. Nearest-rank complete-request wall p50/p95 **4.575/4.725 s**, excluding **140.31 s** startup. Loaded policy-plus-DirectML process-tree RSS was 11.69 GB, not a loading peak. | [report](profile20.json), [driver](profile20_driver.log), [worker](profile20_worker.log), [actions](profile20_actions.npy) |
| Native Windows CPU comparison | Separate [CPU baseline](../internvla_omni_windows_cpu/README.md) was 4.878/4.902 s p50/p95, so this warmed serial hybrid run was 0.303/0.178 s lower at those ranks; startup was 4.72 s longer. Action max absolute difference `0.09171`, relative L2 `2.4391%`, cosine `0.999705`. These are separate runs on one synthetic fixture, not paired randomized measurements or a task-quality pass. | [hybrid profile](profile20.json), [CPU profile](../internvla_omni_windows_cpu/cpu_profile20.json) |
| Public API | `AsyncOmni.generate` returned one finite terminal `action` event with the same action hash. Its startup was 170.50 s and request wall 4.374 s; OS-available RAM before launch was 22.96 GB. | [report](public.json), [driver](public_driver.log), [worker](public_worker.log) |
| Admission and cancellation | A declared 16 GiB host-RAM capacity/demand admitted the profile; an 8 GiB demand was refused before worker launch. In-flight abort produced no stale output; the reservation and quarantine counts were zero after abort and shutdown. | [profile](profile20.json), [refusal](refusal.json), [refusal log](refusal_driver.log) |
The [first complete-action smoke](smoke.json) passed but is one cold sample,
not a latency distribution. Native Windows uses the same action contract as
the [WSL CPU+Radeon path](../internvla_omni_policy/README.md), but the two
runtime/OS profiles are separate experiments. No real Place_Markpen
observation/reference-action tolerance, physical control metadata,
post-cancel restart, concurrent admission, loading peak or sustained
power/thermal behavior was verified. The small warm latency difference and
precision change do not yet justify making the hybrid placement a default.
Reproduce from the repository root in PowerShell with the pinned local files:
```powershell
$root = '\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer'
$repo = Join-Path $root 'vllm-omni-edge'
$evidence = Join-Path $repo 'benchmarks\edge_harness\results\e2e_expansion_20260923\evidence\internvla_omni_windows_hybrid'
$py = 'C:\Users\zhout\w2\omni029venv\Scripts\python.exe'
$env:PYTHONUTF8 = '1'
$env:PYTHONPATH = $repo
$env:VLLM_ENABLE_V1_MULTIPROCESSING = '0'
$env:VLLM_OMNI_EXTERNAL_PYTHON_TORCH_DML = 'C:\Users\zhout\w2\qwen_dml_venv\Scripts\python.exe'
$common = @(
'--model-dir', (Join-Path $root 'models\InternVLA-A1-3B-FT-Place_Markpen'),
'--processor-dir', (Join-Path $root 'models\Qwen3-VL-2B-Instruct-processor'),
'--cosmos-dir', (Join-Path $root 'models\Cosmos-Tokenizer-CI8x8'),
'--graph-file', (Join-Path $root 'models\InternVLA-cosmos-encoder-batch6.pt2'),
'--python-bin', $py, '--placement', 'radeon-cosmos',
'--capacity-gib', '16', '--reserve-gib', '16'
)
& $py (Join-Path $repo 'benchmarks\edge_harness\experiments\probe_omni_internvla_policy.py') @common `
--warmups 1 --repeats 20 --abort-check `
--reference-actions (Join-Path $repo 'benchmarks\edge_harness\results\e2e_expansion_20260923\evidence\internvla_omni_windows_cpu\cpu_profile20_actions.npy') `
--log-file (Join-Path $evidence 'profile20_worker.log') `
--output-report (Join-Path $evidence 'profile20.json') `
--output-actions (Join-Path $evidence 'profile20_actions.npy')
& $py (Join-Path $repo 'benchmarks\edge_harness\experiments\probe_omni_internvla_entrypoint.py') @common `
--log-file (Join-Path $evidence 'public_worker.log') `
--output-report (Join-Path $evidence 'public.json')
```
Next gate: real observations/reference actions and a task tolerance, physical
action metadata and deadlines, then post-cancel restart, peak shared memory,
concurrent load and sustained tail/power/thermal measurement.
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/profile20.json
+1407
View File
File diff suppressed because it is too large. Load diff
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/profile20_actions.npy
BIN
View File
Binary file not shown.
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/profile20_driver.log
+409
View File
@@ -0,0 +1,409 @@
INFO 09-24 02:10:59 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:10:59 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:11:40 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
WARNING 09-24 02:11:40 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 13] Permission denied: 'C:\\Users\\zhout'
{
"scope": "real InternVLA Place_Markpen checkpoint via bounded Omni whole-policy graph stage; synthetic patterned observations/noise; no robot-task quality claim",
"placement": "radeon-cosmos",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"budget": {
"capacities": {
"host_ram": 17179869184
},
"demands": {
"host_ram": 17179869184
}
},
"host_available_bytes_before": 23984087040,
"warmups": 1,
"repeats": 20,
"startup_s": 140.31401089997962,
"execution_plan_start": {
"backend": "external.internvla.policy.v1",
"stage_id": 0,
"worker_generation": "84b2e0bcea424d61bf85403f6ff04dd4",
"worker_pid": 26448,
"worker_script_sha256": "55ffe4ce194eee7e330a4e5def8a2bfec0bfbb98b182885e094f71ea57e4ac9a",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"placement": "radeon-cosmos",
"worker_props": {
"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen",
"processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor",
"cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8",
"placement": "radeon-cosmos",
"torch": "2.13.0+cu130",
"policy_dtype": "bfloat16",
"cosmos_dtype": "float32",
"policy_device": "cpu",
"runtime_mode": "real_checkpoint_loaded",
"policy_load_s": 74.39593969995622,
"action_shape": [
1,
50,
32
],
"image_shape": [
1,
2,
3,
224,
224
],
"state_shape": [
1,
32
],
"action_mode": "delta",
"action_units": "unverified",
"action_joint_order": "unverified",
"action_step_s": null,
"control_ready": false,
"external_load": {
"ep": "dml",
"graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2",
"device_name": "AMD Radeon(TM) 890M Graphics",
"placement_granularity": "output_device",
"session_providers": [
"torch-directml"
],
"available_providers": [
"torch-directml"
],
"node_counts": {
"dml": 1,
"cpu": 0
},
"total_nodes": 1,
"target_nodes": 1,
"fraction_on_target": 1.0,
"session_create_s": 2.0201094000367448,
"warmup_s": 0.8657134999521077,
"rss_bytes": 0,
"inputs": [
{
"name": "pixels",
"shape": [
6,
3,
256,
256
]
}
],
"outputs": [
{
"name": "out_0",
"shape": [
6,
16,
32,
32
]
}
],
"profile_path": "",
"ep_directory": "",
"requested_provider_missing": "",
"fallback_note": "",
"onnxruntime": "",
"note": ""
}
},
"loaded_rss_bytes": 11690680320,
"loaded_process_tree": [
{
"pid": 26448,
"rss_bytes": 4395008,
"private_bytes": 798720
},
{
"pid": 50276,
"rss_bytes": 7458816,
"private_bytes": 1232896
},
{
"pid": 64696,
"rss_bytes": 8929906688,
"private_bytes": 11091988480
},
{
"pid": 50648,
"rss_bytes": 4415488,
"private_bytes": 843776
},
{
"pid": 17480,
"rss_bytes": 7487488,
"private_bytes": 1269760
},
{
"pid": 54588,
"rss_bytes": 2737016832,
"private_bytes": 3113463808
}
],
"reserved_bytes": {
"host_ram": 17179869184
},
"memory_overhead_bytes": 8589934592,
"max_input_bytes": 8388608,
"max_action_bytes": 1048576,
"request_capacity": 1,
"stateful_session": "InternVLA policy worker-owned; complete action chunk per request",
"evidence": "B"
},
"warmup_results": [
{
"wall_s": 4.308969200006686,
"action_sha256": "c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637",
"stage_event": {
"request_id": "internvla-warmup-0",
"stage_id": 0,
"epoch": 1,
"seq": 1,
"kind": "action",
"worker_generation": "84b2e0bcea424d61bf85403f6ff04dd4",
"buffers": [
{
"object_id": "actions",
"owner": "0",
"generation": "84b2e0bcea424d61bf85403f6ff04dd4",
"dtype": "float32",
"shape": [
1,
50,
32
],
"nbytes": 6400,
"memory_domain": "host-copy",
"layout": "C",
"ready": true
}
],
"terminal": true,
"error": null
},
"action_metadata": {
"observation_timestamp_ns": 1790187240075975000,
"generation_timestamp_ns": 1790187244379567500,
"observation_age_s": 4.3035925,
"action_mode": "delta",
"action_shape": [
1,
50,
32
],
"action_units": "unverified",
"joint_order": "unverified",
"action_step_s": null,
"control_ready": false
},
"metrics": {
"policy_wall_s": 4.301543799927458,
"worker_wall_s": 4.273181899916381
}
}
],
"nearest_rank_p50_wall_s": 4.574687099899165,
"nearest_rank_p95_wall_s": 4.7248431000625715,
"unique_action_hashes": [
"c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637"
],
"output_actions": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260923\\evidence\\internvla_omni_windows_hybrid\\profile20_actions.npy",
"reference_comparison": {
"max_abs": 0.09171417355537415,
"relative_l2": 0.024391021579504013,
"cosine": 0.9997047185897827
},
"execution_plan_end": {
"backend": "external.internvla.policy.v1",
"stage_id": 0,
"worker_generation": "84b2e0bcea424d61bf85403f6ff04dd4",
"worker_pid": 26448,
"worker_script_sha256": "55ffe4ce194eee7e330a4e5def8a2bfec0bfbb98b182885e094f71ea57e4ac9a",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"placement": "radeon-cosmos",
"worker_props": {
"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen",
"processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor",
"cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8",
"placement": "radeon-cosmos",
"torch": "2.13.0+cu130",
"policy_dtype": "bfloat16",
"cosmos_dtype": "float32",
"policy_device": "cpu",
"runtime_mode": "real_checkpoint_loaded",
"policy_load_s": 74.39593969995622,
"action_shape": [
1,
50,
32
],
"image_shape": [
1,
2,
3,
224,
224
],
"state_shape": [
1,
32
],
"action_mode": "delta",
"action_units": "unverified",
"action_joint_order": "unverified",
"action_step_s": null,
"control_ready": false,
"external_load": {
"ep": "dml",
"graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2",
"device_name": "AMD Radeon(TM) 890M Graphics",
"placement_granularity": "output_device",
"session_providers": [
"torch-directml"
],
"available_providers": [
"torch-directml"
],
"node_counts": {
"dml": 1,
"cpu": 0
},
"total_nodes": 1,
"target_nodes": 1,
"fraction_on_target": 1.0,
"session_create_s": 2.0201094000367448,
"warmup_s": 0.8657134999521077,
"rss_bytes": 0,
"inputs": [
{
"name": "pixels",
"shape": [
6,
3,
256,
256
]
}
],
"outputs": [
{
"name": "out_0",
"shape": [
6,
16,
32,
32
]
}
],
"profile_path": "",
"ep_directory": "",
"requested_provider_missing": "",
"fallback_note": "",
"onnxruntime": "",
"note": ""
}
},
"loaded_rss_bytes": 11690680320,
"loaded_process_tree": [
{
"pid": 26448,
"rss_bytes": 4395008,
"private_bytes": 798720
},
{
"pid": 50276,
"rss_bytes": 7458816,
"private_bytes": 1232896
},
{
"pid": 64696,
"rss_bytes": 8929906688,
"private_bytes": 11091988480
},
{
"pid": 50648,
"rss_bytes": 4415488,
"private_bytes": 843776
},
{
"pid": 17480,
"rss_bytes": 7487488,
"private_bytes": 1269760
},
{
"pid": 54588,
"rss_bytes": 2737016832,
"private_bytes": 3113463808
}
],
"reserved_bytes": {
"host_ram": 17179869184
},
"memory_overhead_bytes": 8589934592,
"max_input_bytes": 8388608,
"max_action_bytes": 1048576,
"request_capacity": 1,
"stateful_session": "InternVLA policy worker-owned; complete action chunk per request",
"evidence": "B"
},
"abort_check": {
"stale_output": false,
"ledger": {
"capacities": {
"host_ram": 17179869184
},
"reserved": {
"host_ram": 0
},
"owners": [],
"quarantined": []
}
},
"status": "passed",
"ledger_after_shutdown": {
"capacities": {
"host_ram": 17179869184
},
"reserved": {
"host_ram": 0
},
"owners": [],
"quarantined": []
}
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/profile20_worker.log
+28
View File
@@ -0,0 +1,28 @@
W0924 02:12:30.765000 64696 Lib\site-packages\torch\utils\cpp_extension.py:178] No CUDA runtime is found, using CUDA_HOME='C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4'
INFO 09-24 02:12:33 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:12:33 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:12:38 [pipeline_internvla_a1.py:164] Loading InternVLA-A1 weights from \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen
internvla-worker ready port=63927 props={"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen", "processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor", "cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8", "placement": "radeon-cosmos", "torch": "2.13.0+cu130", "policy_dtype": "bfloat16", "cosmos_dtype": "float32", "policy_device": "cpu", "runtime_mode": "real_checkpoint_loaded", "policy_load_s": 74.39593969995622, "action_shape": [1, 50, 32], "image_shape": [1, 2, 3, 224, 224], "state_shape": [1, 32], "action_mode": "delta", "action_units": "unverified", "action_joint_order": "unverified", "action_step_s": null, "control_ready": false, "external_load": {"ep": "dml", "graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2", "device_name": "AMD Radeon(TM) 890M Graphics", "placement_granularity": "output_device", "session_providers": ["torch-directml"], "available_providers": ["torch-directml"], "node_counts": {"dml": 1, "cpu": 0}, "total_nodes": 1, "target_nodes": 1, "fraction_on_target": 1.0, "session_create_s": 2.0201094000367448, "warmup_s": 0.8657134999521077, "rss_bytes": 0, "inputs": [{"name": "pixels", "shape": [6, 3, 256, 256]}], "outputs": [{"name": "out_0", "shape": [6, 16, 32, 32]}], "profile_path": "", "ep_directory": "", "requested_provider_missing": "", "fallback_note": "", "onnxruntime": "", "note": ""}}
127.0.0.1 - - [24/Sep/2026 02:13:59] "GET /health HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:13:59] "GET /props HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:04] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:08] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:13] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:18] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:22] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:27] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:31] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:36] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:40] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:45] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:49] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:54] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:14:59] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:03] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:08] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:12] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:16] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:21] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:25] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:30] "POST /v1/actions HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:15:35] "POST /v1/actions HTTP/1.1" 200 -
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/public.json
+67
View File
@@ -0,0 +1,67 @@
{
"entrypoint": "AsyncOmni.generate",
"placement": "radeon-cosmos",
"pipeline": "internvla_a1_external_whole",
"host_available_bytes_before": 22957715456,
"budget": {
"capacity_gib": 16,
"reserve_gib": 16
},
"startup_s": 170.5034694999922,
"request_wall_s": 4.3736909999279305,
"outputs": [
{
"request_id": "public-internvla-1",
"action_shape": [
1,
50,
32
],
"action_sha256": "c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637",
"finite": true,
"metadata": {
"observation_timestamp_ns": 1790187587426321900,
"generation_timestamp_ns": 1790187591783461200,
"observation_age_s": 4.3571393,
"action_mode": "delta",
"action_shape": [
1,
50,
32
],
"action_units": "unverified",
"joint_order": "unverified",
"action_step_s": null,
"control_ready": false
},
"stage_event": {
"request_id": "public-internvla-1-a9964473",
"stage_id": 0,
"epoch": 1,
"seq": 1,
"kind": "action",
"worker_generation": "5c9b9164205940b8ba27bb3eebb0e25f",
"buffers": [
{
"object_id": "actions",
"owner": "0",
"generation": "5c9b9164205940b8ba27bb3eebb0e25f",
"dtype": "float32",
"shape": [
1,
50,
32
],
"nbytes": 6400,
"memory_domain": "host-copy",
"layout": "C",
"ready": true
}
],
"terminal": true,
"error": null
}
}
],
"status": "passed"
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/public_driver.log
+85
View File
@@ -0,0 +1,85 @@
INFO 09-24 02:16:16 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:16:16 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:16:56 [omni_base.py:216] [AsyncOmni] Initializing with model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen
INFO 09-24 02:16:56 [async_omni_engine.py:172] [AsyncOmniEngine] Initializing with model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen
ERROR 09-24 02:16:57 [repo_utils.py:119] Error parsing config for \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen: Unrecognized model in \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen. Should have a `model_type` key in its config.json., retrying 1 of 2
ERROR 09-24 02:16:59 [repo_utils.py:117] Error parsing config for \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen: Unrecognized model in \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen. Should have a `model_type` key in its config.json.
INFO 09-24 02:16:59 [async_omni_engine.py:290] [AsyncOmniEngine] Launching Orchestrator thread with 1 stages
INFO 09-24 02:16:59 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
INFO 09-24 02:16:59 [initialization.py:356] Loaded OmniTransferConfig with 0 connector configurations
INFO 09-24 02:19:47 [orchestrator.py:604] [Orchestrator] Starting event loop
INFO 09-24 02:19:47 [async_omni_engine.py:323] [AsyncOmniEngine] Orchestrator ready with 1 stages
INFO 09-24 02:19:47 [omni_base.py:235] [AsyncOmni] AsyncOmniEngine initialized in 170.50 seconds
INFO 09-24 02:19:47 [omni_base.py:255] [AsyncOmni] Initialized with 1 stages for model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen
INFO 09-24 02:19:51 [omni_base.py:781] [AsyncOmni] Shutting down
INFO 09-24 02:19:51 [async_omni_engine.py:1807] [AsyncOmniEngine] Shutting down Orchestrator
INFO 09-24 02:19:51 [orchestrator.py:705] [Orchestrator] Received shutdown signal
INFO 09-24 02:19:51 [orchestrator.py:3092] [Orchestrator] Shutting down all 1 client(s)
INFO 09-24 02:19:52 [stage_pool.py:1354] [StagePool] Stage 0 replica 0 shut down
{
"entrypoint": "AsyncOmni.generate",
"placement": "radeon-cosmos",
"pipeline": "internvla_a1_external_whole",
"host_available_bytes_before": 22957715456,
"budget": {
"capacity_gib": 16,
"reserve_gib": 16
},
"startup_s": 170.5034694999922,
"request_wall_s": 4.3736909999279305,
"outputs": [
{
"request_id": "public-internvla-1",
"action_shape": [
1,
50,
32
],
"action_sha256": "c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637",
"finite": true,
"metadata": {
"observation_timestamp_ns": 1790187587426321900,
"generation_timestamp_ns": 1790187591783461200,
"observation_age_s": 4.3571393,
"action_mode": "delta",
"action_shape": [
1,
50,
32
],
"action_units": "unverified",
"joint_order": "unverified",
"action_step_s": null,
"control_ready": false
},
"stage_event": {
"request_id": "public-internvla-1-a9964473",
"stage_id": 0,
"epoch": 1,
"seq": 1,
"kind": "action",
"worker_generation": "5c9b9164205940b8ba27bb3eebb0e25f",
"buffers": [
{
"object_id": "actions",
"owner": "0",
"generation": "5c9b9164205940b8ba27bb3eebb0e25f",
"dtype": "float32",
"shape": [
1,
50,
32
],
"nbytes": 6400,
"memory_domain": "host-copy",
"layout": "C",
"ready": true
}
],
"terminal": true,
"error": null
}
}
],
"status": "passed"
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/public_worker.log
+8
View File
@@ -0,0 +1,8 @@
W0924 02:17:49.139000 27896 Lib\site-packages\torch\utils\cpp_extension.py:178] No CUDA runtime is found, using CUDA_HOME='C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4'
INFO 09-24 02:17:52 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:17:52 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:17:57 [pipeline_internvla_a1.py:164] Loading InternVLA-A1 weights from \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen
internvla-worker ready port=64382 props={"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen", "processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor", "cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8", "placement": "radeon-cosmos", "torch": "2.13.0+cu130", "policy_dtype": "bfloat16", "cosmos_dtype": "float32", "policy_device": "cpu", "runtime_mode": "real_checkpoint_loaded", "policy_load_s": 99.18314119998831, "action_shape": [1, 50, 32], "image_shape": [1, 2, 3, 224, 224], "state_shape": [1, 32], "action_mode": "delta", "action_units": "unverified", "action_joint_order": "unverified", "action_step_s": null, "control_ready": false, "external_load": {"ep": "dml", "graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2", "device_name": "AMD Radeon(TM) 890M Graphics", "placement_granularity": "output_device", "session_providers": ["torch-directml"], "available_providers": ["torch-directml"], "node_counts": {"dml": 1, "cpu": 0}, "total_nodes": 1, "target_nodes": 1, "fraction_on_target": 1.0, "session_create_s": 2.2300097999395803, "warmup_s": 0.8993501000804827, "rss_bytes": 0, "inputs": [{"name": "pixels", "shape": [6, 3, 256, 256]}], "outputs": [{"name": "out_0", "shape": [6, 16, 32, 32]}], "profile_path": "", "ep_directory": "", "requested_provider_missing": "", "fallback_note": "", "onnxruntime": "", "note": ""}}
127.0.0.1 - - [24/Sep/2026 02:19:47] "GET /health HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:19:47] "GET /props HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:19:51] "POST /v1/actions HTTP/1.1" 200 -
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/refusal.json
+39
View File
@@ -0,0 +1,39 @@
{
"scope": "real InternVLA Place_Markpen checkpoint via bounded Omni whole-policy graph stage; synthetic patterned observations/noise; no robot-task quality claim",
"placement": "radeon-cosmos",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"budget": {
"capacities": {
"host_ram": 17179869184
},
"demands": {
"host_ram": 8589934592
}
},
"host_available_bytes_before": 30998159360,
"warmups": 0,
"repeats": 1,
"admission_refusal": "ResourceUnavailable: InternVLA weights, graph and state/workspace/headroom exceed reservation",
"status": "passed",
"ledger_after_shutdown": {
"capacities": {
"host_ram": 17179869184
},
"reserved": {
"host_ram": 0
},
"owners": [],
"quarantined": []
}
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/refusal_driver.log
+80
View File
@@ -0,0 +1,80 @@
INFO 09-24 02:20:29 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:20:29 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:21:13 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
WARNING 09-24 02:21:13 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 13] Permission denied: 'C:\\Users\\zhout'
ERROR 09-24 02:21:50 [stage_runtime.py:265] [StageRuntime] Stage initialization failed; shutting down 0 initialized client(s)
ERROR 09-24 02:21:50 [stage_runtime.py:265] Traceback (most recent call last):
ERROR 09-24 02:21:50 [stage_runtime.py:265] File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 252, in initialize
ERROR 09-24 02:21:50 [stage_runtime.py:265] initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
ERROR 09-24 02:21:50 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 09-24 02:21:50 [stage_runtime.py:265] File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 589, in _initialize_stage_replicas
ERROR 09-24 02:21:50 [stage_runtime.py:265] raise primary_exc
ERROR 09-24 02:21:50 [stage_runtime.py:265] File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 550, in _init_group
ERROR 09-24 02:21:50 [stage_runtime.py:265] client = self._initialize_replica(
ERROR 09-24 02:21:50 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 09-24 02:21:50 [stage_runtime.py:265] File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 817, in _initialize_replica
ERROR 09-24 02:21:50 [stage_runtime.py:265] return create_graph_client(
ERROR 09-24 02:21:50 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^
ERROR 09-24 02:21:50 [stage_runtime.py:265] File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\__init__.py", line 31, in create_graph_client
ERROR 09-24 02:21:50 [stage_runtime.py:265] return InternVLAStageClient(metadata, backend, ledger, reservation)
ERROR 09-24 02:21:50 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 09-24 02:21:50 [stage_runtime.py:265] File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\internvla.py", line 111, in __init__
ERROR 09-24 02:21:50 [stage_runtime.py:265] raise ResourceUnavailable("InternVLA weights, graph and state/workspace/headroom exceed reservation")
ERROR 09-24 02:21:50 [stage_runtime.py:265] vllm_omni.engine.resource_ledger.ResourceUnavailable: InternVLA weights, graph and state/workspace/headroom exceed reservation
{
"scope": "real InternVLA Place_Markpen checkpoint via bounded Omni whole-policy graph stage; synthetic patterned observations/noise; no robot-task quality claim",
"placement": "radeon-cosmos",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"budget": {
"capacities": {
"host_ram": 17179869184
},
"demands": {
"host_ram": 8589934592
}
},
"host_available_bytes_before": 30998159360,
"warmups": 0,
"repeats": 1,
"admission_refusal": "ResourceUnavailable: InternVLA weights, graph and state/workspace/headroom exceed reservation",
"status": "passed",
"ledger_after_shutdown": {
"capacities": {
"host_ram": 17179869184
},
"reserved": {
"host_ram": 0
},
"owners": [],
"quarantined": []
}
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/smoke.json
+393
View File
@@ -0,0 +1,393 @@
{
"scope": "real InternVLA Place_Markpen checkpoint via bounded Omni whole-policy graph stage; synthetic patterned observations/noise; no robot-task quality claim",
"placement": "radeon-cosmos",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"budget": {
"capacities": {
"host_ram": 17179869184
},
"demands": {
"host_ram": 17179869184
}
},
"host_available_bytes_before": 31030185984,
"warmups": 0,
"repeats": 1,
"startup_s": 140.35938389995135,
"execution_plan_start": {
"backend": "external.internvla.policy.v1",
"stage_id": 0,
"worker_generation": "d1563cd4f89b46dd955ea99b7a7cb6fd",
"worker_pid": 40992,
"worker_script_sha256": "55ffe4ce194eee7e330a4e5def8a2bfec0bfbb98b182885e094f71ea57e4ac9a",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"placement": "radeon-cosmos",
"worker_props": {
"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen",
"processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor",
"cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8",
"placement": "radeon-cosmos",
"torch": "2.13.0+cu130",
"policy_dtype": "bfloat16",
"cosmos_dtype": "float32",
"policy_device": "cpu",
"runtime_mode": "real_checkpoint_loaded",
"policy_load_s": 73.86743370001204,
"action_shape": [
1,
50,
32
],
"image_shape": [
1,
2,
3,
224,
224
],
"state_shape": [
1,
32
],
"action_mode": "delta",
"action_units": "unverified",
"action_joint_order": "unverified",
"action_step_s": null,
"control_ready": false,
"external_load": {
"ep": "dml",
"graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2",
"device_name": "AMD Radeon(TM) 890M Graphics",
"placement_granularity": "output_device",
"session_providers": [
"torch-directml"
],
"available_providers": [
"torch-directml"
],
"node_counts": {
"dml": 1,
"cpu": 0
},
"total_nodes": 1,
"target_nodes": 1,
"fraction_on_target": 1.0,
"session_create_s": 2.3821033999556676,
"warmup_s": 0.8999756999546662,
"rss_bytes": 0,
"inputs": [
{
"name": "pixels",
"shape": [
6,
3,
256,
256
]
}
],
"outputs": [
{
"name": "out_0",
"shape": [
6,
16,
32,
32
]
}
],
"profile_path": "",
"ep_directory": "",
"requested_provider_missing": "",
"fallback_note": "",
"onnxruntime": "",
"note": ""
}
},
"loaded_rss_bytes": 11674959872,
"loaded_process_tree": [
{
"pid": 40992,
"rss_bytes": 4386816,
"private_bytes": 782336
},
{
"pid": 57196,
"rss_bytes": 7471104,
"private_bytes": 1236992
},
{
"pid": 16412,
"rss_bytes": 8914538496,
"private_bytes": 11075821568
},
{
"pid": 48204,
"rss_bytes": 4407296,
"private_bytes": 839680
},
{
"pid": 39152,
"rss_bytes": 7487488,
"private_bytes": 1290240
},
{
"pid": 41736,
"rss_bytes": 2736668672,
"private_bytes": 3110862848
}
],
"reserved_bytes": {
"host_ram": 17179869184
},
"memory_overhead_bytes": 8589934592,
"max_input_bytes": 8388608,
"max_action_bytes": 1048576,
"request_capacity": 1,
"stateful_session": "InternVLA policy worker-owned; complete action chunk per request",
"evidence": "B"
},
"warmup_results": [],
"measured": [
{
"wall_s": 4.266927799908444,
"action_sha256": "c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637",
"stage_event": {
"request_id": "internvla-measured-0",
"stage_id": 0,
"epoch": 1,
"seq": 1,
"kind": "action",
"worker_generation": "d1563cd4f89b46dd955ea99b7a7cb6fd",
"buffers": [
{
"object_id": "actions",
"owner": "0",
"generation": "d1563cd4f89b46dd955ea99b7a7cb6fd",
"dtype": "float32",
"shape": [
1,
50,
32
],
"nbytes": 6400,
"memory_domain": "host-copy",
"layout": "C",
"ready": true
}
],
"terminal": true,
"error": null
},
"action_metadata": {
"observation_timestamp_ns": 1790187016402806300,
"generation_timestamp_ns": 1790187020664813100,
"observation_age_s": 4.2620068,
"action_mode": "delta",
"action_shape": [
1,
50,
32
],
"action_units": "unverified",
"joint_order": "unverified",
"action_step_s": null,
"control_ready": false
},
"metrics": {
"policy_wall_s": 4.262615300016478,
"worker_wall_s": 4.22401040000841
}
}
],
"nearest_rank_p50_wall_s": 4.266927799908444,
"nearest_rank_p95_wall_s": 4.266927799908444,
"unique_action_hashes": [
"c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637"
],
"output_actions": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260923\\evidence\\internvla_omni_windows_hybrid\\smoke_actions.npy",
"reference_comparison": {
"max_abs": 0.09171417355537415,
"relative_l2": 0.024391021579504013,
"cosine": 0.9997047185897827
},
"execution_plan_end": {
"backend": "external.internvla.policy.v1",
"stage_id": 0,
"worker_generation": "d1563cd4f89b46dd955ea99b7a7cb6fd",
"worker_pid": 40992,
"worker_script_sha256": "55ffe4ce194eee7e330a4e5def8a2bfec0bfbb98b182885e094f71ea57e4ac9a",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"placement": "radeon-cosmos",
"worker_props": {
"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen",
"processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor",
"cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8",
"placement": "radeon-cosmos",
"torch": "2.13.0+cu130",
"policy_dtype": "bfloat16",
"cosmos_dtype": "float32",
"policy_device": "cpu",
"runtime_mode": "real_checkpoint_loaded",
"policy_load_s": 73.86743370001204,
"action_shape": [
1,
50,
32
],
"image_shape": [
1,
2,
3,
224,
224
],
"state_shape": [
1,
32
],
"action_mode": "delta",
"action_units": "unverified",
"action_joint_order": "unverified",
"action_step_s": null,
"control_ready": false,
"external_load": {
"ep": "dml",
"graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2",
"device_name": "AMD Radeon(TM) 890M Graphics",
"placement_granularity": "output_device",
"session_providers": [
"torch-directml"
],
"available_providers": [
"torch-directml"
],
"node_counts": {
"dml": 1,
"cpu": 0
},
"total_nodes": 1,
"target_nodes": 1,
"fraction_on_target": 1.0,
"session_create_s": 2.3821033999556676,
"warmup_s": 0.8999756999546662,
"rss_bytes": 0,
"inputs": [
{
"name": "pixels",
"shape": [
6,
3,
256,
256
]
}
],
"outputs": [
{
"name": "out_0",
"shape": [
6,
16,
32,
32
]
}
],
"profile_path": "",
"ep_directory": "",
"requested_provider_missing": "",
"fallback_note": "",
"onnxruntime": "",
"note": ""
}
},
"loaded_rss_bytes": 11674959872,
"loaded_process_tree": [
{
"pid": 40992,
"rss_bytes": 4386816,
"private_bytes": 782336
},
{
"pid": 57196,
"rss_bytes": 7471104,
"private_bytes": 1236992
},
{
"pid": 16412,
"rss_bytes": 8914538496,
"private_bytes": 11075821568
},
{
"pid": 48204,
"rss_bytes": 4407296,
"private_bytes": 839680
},
{
"pid": 39152,
"rss_bytes": 7487488,
"private_bytes": 1290240
},
{
"pid": 41736,
"rss_bytes": 2736668672,
"private_bytes": 3110862848
}
],
"reserved_bytes": {
"host_ram": 17179869184
},
"memory_overhead_bytes": 8589934592,
"max_input_bytes": 8388608,
"max_action_bytes": 1048576,
"request_capacity": 1,
"stateful_session": "InternVLA policy worker-owned; complete action chunk per request",
"evidence": "B"
},
"status": "passed",
"ledger_after_shutdown": {
"capacities": {
"host_ram": 17179869184
},
"reserved": {
"host_ram": 0
},
"owners": [],
"quarantined": []
}
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/smoke_actions.npy
BIN
View File
Binary file not shown.
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/smoke_driver.log
+345
View File
@@ -0,0 +1,345 @@
INFO 09-24 02:07:11 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:07:12 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:07:56 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
WARNING 09-24 02:07:56 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 13] Permission denied: 'C:\\Users\\zhout'
{
"scope": "real InternVLA Place_Markpen checkpoint via bounded Omni whole-policy graph stage; synthetic patterned observations/noise; no robot-task quality claim",
"placement": "radeon-cosmos",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"budget": {
"capacities": {
"host_ram": 17179869184
},
"demands": {
"host_ram": 17179869184
}
},
"host_available_bytes_before": 31030185984,
"warmups": 0,
"repeats": 1,
"startup_s": 140.35938389995135,
"execution_plan_start": {
"backend": "external.internvla.policy.v1",
"stage_id": 0,
"worker_generation": "d1563cd4f89b46dd955ea99b7a7cb6fd",
"worker_pid": 40992,
"worker_script_sha256": "55ffe4ce194eee7e330a4e5def8a2bfec0bfbb98b182885e094f71ea57e4ac9a",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"placement": "radeon-cosmos",
"worker_props": {
"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen",
"processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor",
"cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8",
"placement": "radeon-cosmos",
"torch": "2.13.0+cu130",
"policy_dtype": "bfloat16",
"cosmos_dtype": "float32",
"policy_device": "cpu",
"runtime_mode": "real_checkpoint_loaded",
"policy_load_s": 73.86743370001204,
"action_shape": [
1,
50,
32
],
"image_shape": [
1,
2,
3,
224,
224
],
"state_shape": [
1,
32
],
"action_mode": "delta",
"action_units": "unverified",
"action_joint_order": "unverified",
"action_step_s": null,
"control_ready": false,
"external_load": {
"ep": "dml",
"graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2",
"device_name": "AMD Radeon(TM) 890M Graphics",
"placement_granularity": "output_device",
"session_providers": [
"torch-directml"
],
"available_providers": [
"torch-directml"
],
"node_counts": {
"dml": 1,
"cpu": 0
},
"total_nodes": 1,
"target_nodes": 1,
"fraction_on_target": 1.0,
"session_create_s": 2.3821033999556676,
"warmup_s": 0.8999756999546662,
"rss_bytes": 0,
"inputs": [
{
"name": "pixels",
"shape": [
6,
3,
256,
256
]
}
],
"outputs": [
{
"name": "out_0",
"shape": [
6,
16,
32,
32
]
}
],
"profile_path": "",
"ep_directory": "",
"requested_provider_missing": "",
"fallback_note": "",
"onnxruntime": "",
"note": ""
}
},
"loaded_rss_bytes": 11674959872,
"loaded_process_tree": [
{
"pid": 40992,
"rss_bytes": 4386816,
"private_bytes": 782336
},
{
"pid": 57196,
"rss_bytes": 7471104,
"private_bytes": 1236992
},
{
"pid": 16412,
"rss_bytes": 8914538496,
"private_bytes": 11075821568
},
{
"pid": 48204,
"rss_bytes": 4407296,
"private_bytes": 839680
},
{
"pid": 39152,
"rss_bytes": 7487488,
"private_bytes": 1290240
},
{
"pid": 41736,
"rss_bytes": 2736668672,
"private_bytes": 3110862848
}
],
"reserved_bytes": {
"host_ram": 17179869184
},
"memory_overhead_bytes": 8589934592,
"max_input_bytes": 8388608,
"max_action_bytes": 1048576,
"request_capacity": 1,
"stateful_session": "InternVLA policy worker-owned; complete action chunk per request",
"evidence": "B"
},
"warmup_results": [],
"nearest_rank_p50_wall_s": 4.266927799908444,
"nearest_rank_p95_wall_s": 4.266927799908444,
"unique_action_hashes": [
"c6ce59129939fdeecad9304882bddcc427834533df2aba102d69c05a72344637"
],
"output_actions": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260923\\evidence\\internvla_omni_windows_hybrid\\smoke_actions.npy",
"reference_comparison": {
"max_abs": 0.09171417355537415,
"relative_l2": 0.024391021579504013,
"cosine": 0.9997047185897827
},
"execution_plan_end": {
"backend": "external.internvla.policy.v1",
"stage_id": 0,
"worker_generation": "d1563cd4f89b46dd955ea99b7a7cb6fd",
"worker_pid": 40992,
"worker_script_sha256": "55ffe4ce194eee7e330a4e5def8a2bfec0bfbb98b182885e094f71ea57e4ac9a",
"artifact_sha256": {
"python": "0b471133e110cfb53a061cad528ce8e517d7b9ac41a0a396c39ad795a487fc14",
"model": "586799539888c69c35f84d083e559cc5f0ba887801d000e70e65f58d3f85c313",
"model_config": "1272ca8fb93cb42f48b039bb298688b2bd50a66f4faa0afcd9e538e9762f48aa",
"train_config": "e50ec42e40ba1925983e0e8db36558d3738bea64e1698045ff7cc08a130e1053",
"stats": "354710650bb91fc0e84a461987b3554b30bef9d5de47567480624484cdfbf3a2",
"cosmos_encoder": "520a3b878e72ee872faccb1211b5d622d194c0fdd177f2f6df097586b8fa98e5",
"cosmos_decoder": "3b7791e6c8f6f92caacddf48ebb314f044bcb8302d0c7f47501d384a7a80f738",
"processor_tokenizer": "a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7",
"processor_config": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
"graph": "9111b3841893acaaa11cf45715d246da3ac9cdd03c7b33fb3c7c20fe38d74704"
},
"placement": "radeon-cosmos",
"worker_props": {
"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen",
"processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor",
"cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8",
"placement": "radeon-cosmos",
"torch": "2.13.0+cu130",
"policy_dtype": "bfloat16",
"cosmos_dtype": "float32",
"policy_device": "cpu",
"runtime_mode": "real_checkpoint_loaded",
"policy_load_s": 73.86743370001204,
"action_shape": [
1,
50,
32
],
"image_shape": [
1,
2,
3,
224,
224
],
"state_shape": [
1,
32
],
"action_mode": "delta",
"action_units": "unverified",
"action_joint_order": "unverified",
"action_step_s": null,
"control_ready": false,
"external_load": {
"ep": "dml",
"graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2",
"device_name": "AMD Radeon(TM) 890M Graphics",
"placement_granularity": "output_device",
"session_providers": [
"torch-directml"
],
"available_providers": [
"torch-directml"
],
"node_counts": {
"dml": 1,
"cpu": 0
},
"total_nodes": 1,
"target_nodes": 1,
"fraction_on_target": 1.0,
"session_create_s": 2.3821033999556676,
"warmup_s": 0.8999756999546662,
"rss_bytes": 0,
"inputs": [
{
"name": "pixels",
"shape": [
6,
3,
256,
256
]
}
],
"outputs": [
{
"name": "out_0",
"shape": [
6,
16,
32,
32
]
}
],
"profile_path": "",
"ep_directory": "",
"requested_provider_missing": "",
"fallback_note": "",
"onnxruntime": "",
"note": ""
}
},
"loaded_rss_bytes": 11674959872,
"loaded_process_tree": [
{
"pid": 40992,
"rss_bytes": 4386816,
"private_bytes": 782336
},
{
"pid": 57196,
"rss_bytes": 7471104,
"private_bytes": 1236992
},
{
"pid": 16412,
"rss_bytes": 8914538496,
"private_bytes": 11075821568
},
{
"pid": 48204,
"rss_bytes": 4407296,
"private_bytes": 839680
},
{
"pid": 39152,
"rss_bytes": 7487488,
"private_bytes": 1290240
},
{
"pid": 41736,
"rss_bytes": 2736668672,
"private_bytes": 3110862848
}
],
"reserved_bytes": {
"host_ram": 17179869184
},
"memory_overhead_bytes": 8589934592,
"max_input_bytes": 8388608,
"max_action_bytes": 1048576,
"request_capacity": 1,
"stateful_session": "InternVLA policy worker-owned; complete action chunk per request",
"evidence": "B"
},
"status": "passed",
"ledger_after_shutdown": {
"capacities": {
"host_ram": 17179869184
},
"reserved": {
"host_ram": 0
},
"owners": [],
"quarantined": []
}
}
benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/smoke_worker.log
+8
View File
@@ -0,0 +1,8 @@
W0924 02:08:46.503000 16412 Lib\site-packages\torch\utils\cpp_extension.py:178] No CUDA runtime is found, using CUDA_HOME='C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4'
INFO 09-24 02:08:49 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
INFO 09-24 02:08:49 [patch.py:510] inductor factorable-divisibility patch: installed.
INFO 09-24 02:08:54 [pipeline_internvla_a1.py:164] Loading InternVLA-A1 weights from \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\InternVLA-A1-3B-FT-Place_Markpen
internvla-worker ready port=52681 props={"model_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-A1-3B-FT-Place_Markpen", "processor_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Qwen3-VL-2B-Instruct-processor", "cosmos_dir": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\Cosmos-Tokenizer-CI8x8", "placement": "radeon-cosmos", "torch": "2.13.0+cu130", "policy_dtype": "bfloat16", "cosmos_dtype": "float32", "policy_device": "cpu", "runtime_mode": "real_checkpoint_loaded", "policy_load_s": 73.86743370001204, "action_shape": [1, 50, 32], "image_shape": [1, 2, 3, 224, 224], "state_shape": [1, 32], "action_mode": "delta", "action_units": "unverified", "action_joint_order": "unverified", "action_step_s": null, "control_ready": false, "external_load": {"ep": "dml", "graph_path": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\InternVLA-cosmos-encoder-batch6.pt2", "device_name": "AMD Radeon(TM) 890M Graphics", "placement_granularity": "output_device", "session_providers": ["torch-directml"], "available_providers": ["torch-directml"], "node_counts": {"dml": 1, "cpu": 0}, "total_nodes": 1, "target_nodes": 1, "fraction_on_target": 1.0, "session_create_s": 2.3821033999556676, "warmup_s": 0.8999756999546662, "rss_bytes": 0, "inputs": [{"name": "pixels", "shape": [6, 3, 256, 256]}], "outputs": [{"name": "out_0", "shape": [6, 16, 32, 32]}], "profile_path": "", "ep_directory": "", "requested_provider_missing": "", "fallback_note": "", "onnxruntime": "", "note": ""}}
127.0.0.1 - - [24/Sep/2026 02:10:16] "GET /health HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:10:16] "GET /props HTTP/1.1" 200 -
127.0.0.1 - - [24/Sep/2026 02:10:20] "POST /v1/actions HTTP/1.1" 200 -
benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
+3 -3
View File
@@ -2,7 +2,7 @@
All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the [2026-09-23 scoped recovery runs](../../../e2e_recovery_20260923/README.md) and [WSL CPU/native Windows expansion](../../../e2e_expansion_20260923/README.md). The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, a Spark output-head component profile on the HX370 AMD NPU, complete Spark text-request profiles through standalone llama.cpp and a bounded Omni whole-session stage on native Windows CPU and Radeon 890M, native Windows CPU Qwen3-TTS text-to-WAV profiles through both standalone PyTorch and a bounded Omni whole-session stage, and a hybrid CPU+Radeon 890M Qwen3-TTS profile likewise repeated through a bounded Omni whole-session stage. A pinned GGUF MiniCPM-o route adds combined image+audio to text+speech runs on WSL CPU, native Windows CPU and native Windows Radeon 890M+CPU. All three now pass a bounded complete request through an Omni graph stage; the WSL CPU path additionally passed one warmup plus 20 serial measured requests. These are scoped synthetic cases, not full release qualification.
The WSL CPU and CPU+Radeon InternVLA cells also have [bounded Omni complete-policy profiles](../../../e2e_expansion_20260923/evidence/internvla_omni_policy/README.md): one warmup and 20 serial patterned synthetic requests per placement, plus admission refusal, cancellation and a public hybrid `AsyncOmni.generate` request. A separate [native Windows CPU Omni profile](../../../e2e_expansion_20260923/evidence/internvla_omni_windows_cpu/README.md) adds 20 serial patterned requests, refusal, cancellation and a public action request. These advance only the synthetic action-policy path. Real observations/reference actions, physical action units/order/step time, task quality, loading peak and sustained power/thermal behavior remain unverified. The earlier [Cosmos component and direct hybrid results](../../../e2e_expansion_20260923/evidence/internvla_cosmos_radeon890m/README.md) remain separate experiments.
The WSL CPU and CPU+Radeon InternVLA cells also have [bounded Omni complete-policy profiles](../../../e2e_expansion_20260923/evidence/internvla_omni_policy/README.md): one warmup and 20 serial patterned synthetic requests per placement, plus admission refusal, cancellation and a public hybrid `AsyncOmni.generate` request. Separate native Windows [CPU](../../../e2e_expansion_20260923/evidence/internvla_omni_windows_cpu/README.md) and [CPU+Radeon](../../../e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/README.md) Omni profiles add 20 serial patterned requests per placement, refusal, cancellation and public action requests. These advance only the synthetic action-policy path. Real observations/reference actions, physical action units/order/step time, task quality, loading peak and sustained power/thermal behavior remain unverified. The earlier [Cosmos component and direct hybrid results](../../../e2e_expansion_20260923/evidence/internvla_cosmos_radeon890m/README.md) remain separate experiments.
## Configuration × model
@@ -20,7 +20,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
| PC HX370: CPU / native Windows | scoped E2E: Omni Spark text | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF image+audio → text+speech | synthetic policy pass: Omni CPU |
| PC HX370 + RTX 5090 Laptop / WSL | profile completed | profile completed | scoped E2E: text + image | scoped E2E: text/image/audio → speech | synthetic policy pass |
| PC HX370 + RTX 5090 Laptop / Windows | profile completed | profile completed | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass |
| PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Omni Radeon talker/codec + CPU predictor | NOT E2E | scoped E2E: Omni Radeon+CPU image+audio → text+speech | synthetic policy pass: Omni CPU policy + Radeon Cosmos |
| PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Omni Radeon talker/codec + CPU predictor | NOT E2E | scoped E2E: Omni Radeon+CPU image+audio → text+speech | synthetic policy pass: Omni CPU+Radeon Cosmos (WSL/native Windows) |
| PC HX370 + AMD NPU / Windows worker | NOT E2E (output-head component passed) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
| PC Snapdragon X Elite CRD / AI Hub | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
@@ -57,7 +57,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
| PC HX370 + Radeon 890M / Windows worker | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID | The pinned Q8_0 CustomVoice talker, F16 codec and punctuation GGUFs completed two named text-to-WAV requests through an Omni whole-session CrispASR stage on native Windows. The runtime log identifies Radeon 890M Vulkan0 for talker/codec and the 432 MB code predictor pinned to CPU FP32; this is not all-Radeon or NPU execution. Both PCM hashes exactly match the earlier standalone outputs. Whisper tiny.en transcribed the first Omni WAV and both standalone WAVs exactly (WER 0 as an intelligibility proxy). One warmup plus 20 serial complete Omni requests for 2.56 s of audio had identical PCM and nearest-rank wall p50/p95 3.298/3.336 s (RTF 1.29/1.30). A public AsyncOmni audio request passed; cancellation drained the worker without stale output or remaining reservation. The earlier standalone profile was 3.310/3.345 s. The default Vulkan configuration previously fell back to CPU, all-Vulkan ran away to a 360-frame cap, and DirectML failed waveform quality. No playable stream, post-cancel restart, loading peak, broad speech quality or sustained power/thermal run is claimed. | Validate playable streaming, restart/recovery, broader voice quality, shared-RAM/GPU peak, concurrency and sustained real-time behavior before wider qualification; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md), [failed DirectML route](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_directml/README.md). |
| PC HX370 + Radeon 890M / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | FP32 448x448 vision tower: DirectML 85/113 nodes; normalized max error 9.095e-06 versus torch CPU. Historical component path, not full 27B or current graph-stage image-to-text integration. | Recover the hashed tower bundle and qualify its current Omni graph handoff with the language stage. |
| PC HX370 + Radeon 890M / Windows worker | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID | The same pinned GGUF set ran one combined synthetic spoken-question+JPEG → text+speech request through a bounded Omni whole-session graph stage. Logs verify `GGML_VK_VISIBLE_DEVICES=1`, Radeon 890M Vulkan0 language (37/37 layers) and vision, with CPU TTS/Token2Wav/vocoder. It answered “The square in the image is red.” and produced 2.60 s of 24 kHz speech in 18.27 s complete-request wall time (one cold sample, zero warmups); sampled worker RSS peaked at 15.57 GB. The stage verified artifact and binary hashes, reserved 24 GiB shared RAM, emitted a terminal audio event, and in-flight cancellation left no stale output or reservation. The prior standalone prompted run took 18.63 s and its output transcribed with WER 0. This is CPU+iGPU joint execution, not all-Radeon or NPU execution; neither run qualifies general quality or sustained performance. | Validate real multimodal quality, playable streaming, post-cancel restart, loading peak, repeated/concurrent latency and sustained thermal behavior. [Omni raw evidence](../../../e2e_expansion_20260923/evidence/minicpmo_omni_cpp/README.md), [standalone prompted evidence](../../../e2e_expansion_20260923/evidence/minicpmo_cpp_radeon890m_prompted/README.md). |
| PC HX370 + Radeon 890M / Windows worker | InternVLA-A1 | SYNTHETIC_POLICY_PASS_HYBRID_OMNI; NOT_TASK_E2E | The real Place_Markpen checkpoint ran in an admitted Omni whole-policy stage: BF16 Qwen3-VL/flow action policy on WSL CPU and the pinned six-frame FP32 Cosmos encoder in Omni's external DirectML worker, with selected Radeon 890M and output-device placement verified. One warmup plus 20 serial patterned synthetic camera/state/noise requests returned identical finite `[1,50,32]` action chunks; nearest-rank complete-request wall p50/p95 was 3.120/3.197 s versus a separately run Omni CPU baseline of 3.175/3.219 s. The hybrid startup was 43.36 s versus CPU 40.03 s. Hybrid-vs-CPU action relative L2 difference was 2.305%, cosine 0.999737; no task tolerance exists. The FP32 encoder is an explicit precision change after DirectML BF16 aborted. The stage verified hashes, reserved 16 GiB shared host RAM, refused an insufficient reservation on the CPU route, emitted a terminal action event through public `AsyncOmni.generate`, and released its reservation after in-flight cancellation and shutdown. An initial hybrid shutdown quarantined 16 GiB; a lifecycle fix and reruns cleared it. This is output-device evidence, not internal per-op attribution or robot-task quality. The older direct hybrid's zero-input 2.952/3.322 s profile is a distinct fixture/runtime and cannot be compared as the same run. | Validate real observation/reference-action quality, physical units/order/step time, post-cancel restart, loading peak, concurrent admission and sustained tail/power before deployment; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/internvla_omni_policy/README.md), [earlier direct hybrid/component evidence](../../../e2e_expansion_20260923/evidence/internvla_cosmos_radeon890m/README.md). |
| PC HX370 + Radeon 890M / Windows worker | InternVLA-A1 | SYNTHETIC_POLICY_PASS_HYBRID_OMNI; NOT_TASK_E2E | The real Place_Markpen checkpoint ran in admitted Omni whole-policy stages on both WSL CPU+Radeon and native Windows CPU+Radeon. In each, the BF16 Qwen3-VL/flow action policy stayed on CPU and a pinned six-frame FP32 Cosmos encoder used Omni's external DirectML worker; the selected Radeon 890M and graph output device were verified, without internal per-op attribution. Each placement returned 20/20 identical finite `[1,50,32]` actions after one warmup. WSL hybrid complete-request p50/p95 was 3.120/3.197 s versus its separate CPU baseline 3.175/3.219 s, with action relative L2 difference 2.305%. Native Windows hybrid was 4.575/4.725 s versus its separate CPU baseline 4.878/4.902 s, with action relative L2 difference 2.4391%. Native hybrid startup was 140.31 s versus CPU 135.60 s. These are separate one-fixture serial runs, not paired speedup or task-quality proof. FP32 encoder is an explicit precision change after DirectML BF16 aborted. Both stages verified hashes, reserved 16 GiB shared host RAM, rejected insufficient demand before launch, emitted terminal actions through public `AsyncOmni.generate`, and released reservations after in-flight cancellation/shutdown. An initial WSL hybrid shutdown quarantined 16 GiB; the lifecycle fix and reruns cleared it. The older direct hybrid zero-input profile is a different experiment. | Validate real observations/reference actions, physical units/order/step time, post-cancel restart, loading peak, concurrency and sustained tail/power before deployment; [WSL Omni evidence](../../../e2e_expansion_20260923/evidence/internvla_omni_policy/README.md), [native Windows Omni evidence](../../../e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/README.md), [earlier direct component evidence](../../../e2e_expansion_20260923/evidence/internvla_cosmos_radeon890m/README.md). |
| PC HX370 + AMD NPU / Windows worker | Spark-X2.5 | COMPONENT_PASS; NOT_E2E_PROFILEABLE | The real 1.7B output head (RMS norm and full 131,072-logit projection) ran through Omni `external.graph.v1` with VitisAI: one fused NPU partition and seven CPU nodes per request. The A16W8 graph matched top-1 on four captured activations (full-logit SNR 55.43–56.81 dB). One warmup plus 20 serial repeated-input component requests gave p50/p95 10.21/12.44 ms, versus 17.97/20.29 ms for the original FP32 output head on the same Omni CPU stage. NPU startup was 18.21 s versus 4.48 s for FP32 CPU, and NPU worker peak RSS was 3.52 GiB versus 2.06 GiB. No full prefill/decode/KV/sampling, real handoff, power, or sustained run was verified. | Measure an actual vLLM-to-output-head handoff and whole-generation cost, state/cancellation, memory admission and quality before considering a split plan; [component evidence](../../../e2e_expansion_20260923/evidence/spark_amd_npu_output_head/README.md). |
| PC HX370 + AMD NPU / Windows worker | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC HX370 + AMD NPU / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The tested A16W8 whole-vision-tower export failed NPU placement. This conclusion is specific to that export, not all future Qwen/NPU combinations. | Validate a supported export or smaller useful component, including same-checkpoint numerical quality. |
docs/design/omni_edge_e2e_roadmap.md
+12 -1
View File
@@ -65,6 +65,17 @@ in-flight abort passed. Native-vs-WSL CPU action relative L2 difference on
the same fixture was 2.1795%, without a task tolerance. This deepens the
native Windows CPU cell, not any robot-task or mobile qualification.
The [native Windows CPU+Radeon Omni stage](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/internvla_omni_windows_hybrid/README.md)
then ran the same fixture with a pinned FP32 six-frame Cosmos graph in the
native DirectML worker and the BF16 policy on CPU. Twenty of twenty requests
returned identical finite actions; complete-request p50/p95 was
4.575/4.725 s versus the separate native CPU run's 4.878/4.902 s. Its
public action output, 8 GiB pre-load refusal and in-flight abort passed
under a 16 GiB shared-RAM reservation. The selected graph output device was
Radeon 890M, but per-operator placement was not measured. Action relative
L2 difference was 2.4391% without a task tolerance; the hybrid startup was
4.72 s longer. Real task quality and sustained benefit remain open.
The WSL RTX MiniCPM-o image-to-text+speech path now passes for one 448×448
synthetic red-square image under an explicit 0.61 thinker GPU budget. The
previous 0.58 budget failed the 2,048-token KV admission gate; the successful
@@ -219,7 +230,7 @@ capability and lifecycle implementation; a state-handle type alone is insufficie
| M2 | Qwen3-TTS | Resolve desktop playback and shutdown findings and CPU compatibility; then mobile talker/predictor plus GPU vocoder | Complete PCM and tail, history/chunk correctness, interruption/recovery, quality checks, no post-startup underruns and RTF below 1 in the declared workload |
| M3 | AMD stages | Connect compatible real encoders or other useful coarse stages through current Omni workers; start from the existing 890M vision evidence | Actual node/device placement, numerical and task quality, complete downstream output, shared-memory and handoff costs; reject unhelpful splits |
| M4a | MiniCPM-o 4.5 | Three-stage desktop text+WAV produced coherent text and nonzero WAV on RTX 5090 Laptop under WSL and native Windows, and WSL CPU, using separate constrained plans. One synthetic red-square image and one synthetic tone each passed through the three-stage WSL RTX, native Windows RTX and WSL CPU paths to text and nonzero WAV. RTX text/token IDs matched across OS environments; CPU wording/token IDs differed while remaining plausible. The WSL RTX image plan needed a thinker GPU budget increase from 0.58 to 0.61 for KV admission. WSL multimodal shutdown cleanup remains open, including forced stage-0 termination on CPU. The CPU plan needed vLLM 0.28/0.29 processor compatibility and used swap; 20 serial **text** CPU requests measured p50 21.86 s/p95 22.97 s. Native Windows needed longer startup limits; 20 serial text requests measured p50 18.38 s/p95 18.73 s with host RAM/pagefile pressure. Add video encoder and persistent-state reference before mobile artifacts | Speech intelligibility/alignment, loading-peak admission, concurrency, broad image/audio-understanding suites, video, shutdown lifecycle, then combined streaming and interruption |
| M4b | InternVLA-A1 | Strict-load synthetic policy forwards pass on WSL and native Windows CPU/RTX. Bounded Omni whole-policy stages pass on WSL CPU, WSL CPU+Radeon Cosmos and native Windows CPU with explicit shared-RAM admission, terminal action metadata and in-flight cancellation; public hybrid and native CPU entrypoints pass. This is not robot-task qualification. | Reference action agreement on real observations, physical units/order/step time, observation age and deadlines; post-cancel restart, loading peak, concurrency, power and thermal behavior. |
| M4b | InternVLA-A1 | Strict-load synthetic policy forwards pass on WSL and native Windows CPU/RTX. Bounded Omni whole-policy stages pass on WSL CPU, WSL CPU+Radeon Cosmos, native Windows CPU and native Windows CPU+Radeon Cosmos with explicit shared-RAM admission, terminal action metadata and in-flight cancellation; their public CPU/hybrid entrypoints pass where measured. This is not robot-task qualification. | Reference action agreement on real observations, physical units/order/step time, observation age and deadlines; post-cancel restart, loading peak, concurrency, power and thermal behavior. |
| Independent | Qwen3.8-27B | Public Omni pipeline binding; text, then image/video and longer contexts; same-model CPU/iGPU artifact and explicit offload evaluation | Modality-specific quality, actual loading/runtime memory and measured performance; each precision/context receives its own qualification |
The initial M0 desktop text acceptance remains scoped to its tested checkpoint,