DGX Spark Supercluster — custom AI infrastructure
Collection
Three-node DGX Spark supercluster: 1M-context DeepSeek-V4.1-Flash EXL3 TP3, independent eval results, reproducible serving profile. • 3 items • Updated
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: ArrowNotImplementedError
Message: Cannot write struct type 'environment' with no child field to Parquet. Consider adding a dummy child field.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1876, in _prepare_split_single
num_examples, num_bytes = writer.finalize()
~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 830, in finalize
self.write_rows_on_file()
~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 690, in write_rows_on_file
self._write_table(table)
~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 817, in _write_table
self._build_writer(inferred_schema=pa_table.schema)
~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 861, in _build_writer
self.pa_writer = pq.ParquetWriter(
~~~~~~~~~~~~~~~~^
self.stream,
^^^^^^^^^^^^
...<3 lines>...
**get_parquet_column_options(self._features, self._schema),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/pyarrow/parquet/core.py", line 1115, in __init__
self.writer = _parquet.ParquetWriter(
~~~~~~~~~~~~~~~~~~~~~~^
sink, schema,
^^^^^^^^^^^^^
...<20 lines>...
max_rows_per_page=max_rows_per_page,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
**options)
^^^^^^^^^^
File "pyarrow/_parquet.pyx", line 2430, in pyarrow._parquet.ParquetWriter.__cinit__
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowNotImplementedError: Cannot write struct type 'environment' with no child field to Parquet. Consider adding a dummy child field.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
schema_version int64 | promoted_baseline string | model string | upstream string | generation_defaults dict | default_shape string | shapes dict | container dict | host dict |
|---|---|---|---|---|---|---|---|---|
2 | 2026-10-30-spark-ds41f-1030 | v41-flash-1017 | https://github.com/christopherowen/spark-ds41f.git | {
"temperature": 1,
"top_p": 1
} | tp4-ring | {
"tp4-ring": {
"nodes": 4,
"sizing": {
"max_model_len": 1048576,
"max_num_seqs": 16,
"max_num_batched_tokens": 16384,
"kv_cache_memory_bytes": 11274289152,
"prefix_cache_retention_interval": 8192
},
"fabric": {
"transport": "oneshot-ring4"
},
"environment":... | {
"image": "spark-ds41f:1030",
"docker_run_args": [
"--network=host",
"--ipc=host",
"--gpus=all",
"--device=/dev/infiniband:/dev/infiniband:rwm",
"--cap-add=IPC_LOCK",
"--ulimit=memlock=-1:-1",
"--ulimit=stack=67108864:67108864",
"--ulimit=core=0",
"--security-opt=seccomp=unconfi... | {
"memguard_gib": 3,
"memguard_interval_seconds": 2,
"startup_memguard_gib": 5,
"startup_memguard_interval_seconds": 0.25,
"startup_memguard_wait_seconds": 120,
"kernel": {
"release": "7.0.0-1019-nvidia-64k",
"page_size": 65536,
"memory_saver_required": true
}
} |
Serving DeepSeek-V4.1-Flash-1017 at 1M context on 3× DGX Spark
Reproducible serving profile for running DeepSeek-V4.1-Flash (EXL3) across three NVIDIA DGX Spark nodes at a 1,048,576-token maximum context. This repository contains the machine-readable profile and the operating notes, not the model weights.
Upstream engine: christopherowen/spark-ds41f
(container build 1030). Weights: deepseek-ai/DeepSeek-V4.1-Flash-1017, revision 1bae0e10.
Why this exists
A single DGX Spark cannot hold a 1M-context activation plus its KV cache. Three of them can, if you get four things right:
- Tensor parallel across a dedicated ring. Each node pair is wired with its own direct 200 Gb/s QSFP segment (a triangle, no switch), addressed on a private /24 fabric that is separate from management traffic. NCCL runs RoCE v2 over that fabric only.
- A hard cap on concurrent sequences. 4 is the stable operating point at 1M context. 6 and 8 both crashed during load.
- A small, explicitly-sized KV cache. 2,362,232,012 bytes. Letting vLLM infer it
from
gpu_memory_utilizationproduces a boot that is one request away from OOM. - A memory guard with a grace period. A guard that kills at a fixed headroom threshold without grace produced ten consecutive false kills on a cold load. The working configuration is a 5 GiB startup floor with a 45 s grace, dropping to a 3 GiB steady floor with a 30 s grace.
Profile
Two shapes are defined in configs/serving-profile.json:
tp3-triangle (used) |
tp4-ring (defined, untested here) |
|
|---|---|---|
| Nodes | 3 | 4 |
max_model_len |
1,048,576 | 1,048,576 |
max_num_seqs |
4 | 16 |
max_num_batched_tokens |
6,144 | 16,384 |
kv_cache_memory_bytes |
2,362,232,012 | 11,274,289,152 |
| prefix-cache retention | 512 | 8,192 |
| fabric transport | oneshot-direct |
oneshot-ring4 |
Container runtime arguments that matter: --network=host, --ipc=host,
--device=/dev/infiniband, --cap-add=IPC_LOCK, --ulimit=memlock=-1:-1,
--shm-size=16g, --memory=112g / --memory-swap=112g (no swap headroom — swapping
a model load is worse than failing it), and a pinned --cpuset-cpus set.
Host kernel: 7.0.0-1019-nvidia-64k, 64 KiB pages, with the NVIDIA memory-saver module required (it lets helpers be paused and their memory released during load).
Boot order and pre-flight
Workers first, head last. Before anything is paused or removed, the profile is rendered for all three ranks and validated:
- the render must succeed for every node,
- every render must contain a
docker run, - all three ranks must agree on
max-num-batched-tokens,max-model-lenandgpu-memory-utilization.
Rank disagreement is a hard abort. A bad profile used to abort mid-boot with helper services already paused, which turned a config typo into a four-minute outage.
Also required before boot: at least 100 GiB MemAvailable on every node, and no
previous vllm_dsv41 container still running.
Measured operating figures
| Metric | Value |
|---|---|
| Single-stream prose | 53.9 tok/s |
| Aggregate, 4-request burst | 137.6 tok/s |
| Aggregate, 6-request burst | 123.2 tok/s |
| 443,223-token needle retrieval | 87.1 s |
| Minimum headroom during the long-context test | 6.8 GiB (head node) |
Evaluation results measured on this profile are published separately as
dgx-spark-supercluster-evals.
Rollback discipline
Every profile change keeps its predecessor addressable: the previous profile file and container image stay on the head node with a one-command revert. Serving changes are promoted only after a comparison run, never in place.
Notes and limits
- These figures are from a specific three-node configuration under bounded tests. They are not a capacity guarantee.
- Heavy multimodal generation (image, video, audio) runs in a separate window with inference paused; it does not coexist with the 1M-context profile on this hardware.
- Everything above was measured, not estimated from vendor specifications.
License
Documentation and configuration: CC BY 4.0.
- Downloads last month
- 65