THE SOURCE IN CONTEXT
The operator pads attention-head layouts to make native tensor parallelism three possible, using a custom target and draft loader. The run uses three GB10 nodes linked in a switchless RoCE triangle, FP8 KV cache and a 262,144-token context ceiling. This is a patched serving stack rather than an ordinary unmodified vLLM launch. The report’s 80.41 GiB per-rank weights and 122-second load are source observations, not a capacity guarantee for another build.
How does the three-node path work?
The operator pads attention-head layouts to make native tensor parallelism three possible, using a custom target and draft loader. The run uses three GB10 nodes linked in a switchless RoCE triangle, FP8 KV cache and a 262,144-token context ceiling. This is a patched serving stack rather than an ordinary unmodified vLLM launch. The report’s 80.41 GiB per-rank weights and 122-second load are source observations, not a capacity guarantee for another build.
What changes when the sequence cap is raised?
The greedy 80-item GSM8K sweep at eight parallel requests reports 64.63 aggregate tokens/s with max_num_seqs three and 97.28 with eight, while both report 98.8% accuracy. That comparison changes admission behavior. It should not be mixed with the short 128-token llama-benchy runs or interpreted as a 97-token/s single-user decode rate.
Why keep the sampled run separate?
At the publisher’s temperature 1.0 and top_p 0.95, the parallel-eight quality run reports 97.5% accuracy and 86.46 aggregate tokens/s. Greedy and sampled runs answer different questions. The 80-item corpus is a small task-specific check; even a perfect result on that subset does not prove general equivalence of a patched sharding implementation.
What is still unknown?
The record does not establish a full original weight hash or executed runtime commit. The document also compares older four-node data, which belongs to a separate configuration. Preserve the exact custom patch set, container, topology, cache policy and workload before attempting a reproduction. This site has not run the stack.
Reported figures and their conditions
Selected facts reported by the original author. Each row has its own configuration and measurement boundary. Run AI Here has not reproduced these tests.
| Configuration / workload | Metric | Reported value | Evidence |
|---|---|---|---|
| Short generation, depth 0, C1 | aggregate output | 23.55 tokens/sreported value | Original row ↗ |
| Generation, depth 100000, C1 | aggregate output | 26.83 tokens/sreported value | Original row ↗ |
| GSM8K greedy, parallel 8, sequence cap 3 | aggregate output | 64.63 tokens/sreported value | Original row ↗ |
| GSM8K greedy, parallel 8, sequence cap 8 | aggregate output | 97.28 tokens/sreported value | Original row ↗ |
| GSM8K sampled, parallel 8 | accuracy | 97.5 percentreported value | Original row ↗ |
Short generation, depth 0, C1 · aggregate output · conditions
| system | NVIDIA DGX Spark / GB10 |
|---|---|
| node_count | 3 |
| parallelism | custom native vLLM TP=3 |
| kv_dtype | FP8 |
| context_limit_tokens | 262144 |
| experiment_date | 2026-08-05 |
| model | nvidia/MiniMax-M3-NVFP4 |
| draft | nvidia/MiniMax-M3-DSpark |
| speculative_tokens | 8 |
| runtime | custom vLLM TP3 |
| prefix_cache | True |
| harness | llama-benchy 0.4.0 |
| repetitions | 3 |
| max_num_seqs | 3 |
| output_tokens | 128 |
| retained_depth_tokens | 0 |
| concurrency | 1 |
Generation, depth 100000, C1 · aggregate output · conditions
| system | NVIDIA DGX Spark / GB10 |
|---|---|
| node_count | 3 |
| parallelism | custom native vLLM TP=3 |
| kv_dtype | FP8 |
| context_limit_tokens | 262144 |
| experiment_date | 2026-08-05 |
| model | nvidia/MiniMax-M3-NVFP4 |
| draft | nvidia/MiniMax-M3-DSpark |
| speculative_tokens | 8 |
| runtime | custom vLLM TP3 |
| prefix_cache | True |
| harness | llama-benchy 0.4.0 |
| repetitions | 3 |
| max_num_seqs | 3 |
| output_tokens | 128 |
| retained_depth_tokens | 100000 |
| concurrency | 1 |
GSM8K greedy, parallel 8, sequence cap 3 · aggregate output · conditions
| system | NVIDIA DGX Spark / GB10 |
|---|---|
| node_count | 3 |
| parallelism | custom native vLLM TP=3 |
| kv_dtype | FP8 |
| context_limit_tokens | 262144 |
| experiment_date | 2026-08-05 |
| model | nvidia/MiniMax-M3-NVFP4 |
| draft | nvidia/MiniMax-M3-DSpark |
| speculative_tokens | 8 |
| runtime | custom vLLM TP3 |
| prefix_cache | True |
| harness | llama-benchy 0.4.0 |
| repetitions | Not specified |
| max_num_seqs | 3 |
| corpus_items | 80 |
| temperature | 0 |
| concurrency | 8 |
GSM8K greedy, parallel 8, sequence cap 8 · aggregate output · conditions
| system | NVIDIA DGX Spark / GB10 |
|---|---|
| node_count | 3 |
| parallelism | custom native vLLM TP=3 |
| kv_dtype | FP8 |
| context_limit_tokens | 262144 |
| experiment_date | 2026-08-05 |
| model | nvidia/MiniMax-M3-NVFP4 |
| draft | nvidia/MiniMax-M3-DSpark |
| speculative_tokens | 8 |
| runtime | custom vLLM TP3 |
| prefix_cache | True |
| harness | llama-benchy 0.4.0 |
| repetitions | Not specified |
| max_num_seqs | 8 |
| corpus_items | 80 |
| temperature | 0 |
| concurrency | 8 |
GSM8K sampled, parallel 8 · accuracy · conditions
| system | NVIDIA DGX Spark / GB10 |
|---|---|
| node_count | 3 |
| parallelism | custom native vLLM TP=3 |
| kv_dtype | FP8 |
| context_limit_tokens | 262144 |
| experiment_date | 2026-08-05 |
| model | nvidia/MiniMax-M3-NVFP4 |
| draft | nvidia/MiniMax-M3-DSpark |
| speculative_tokens | 8 |
| runtime | custom vLLM TP3 |
| prefix_cache | True |
| harness | llama-benchy 0.4.0 |
| repetitions | Not specified |
| max_num_seqs | 8 |
| corpus_items | 80 |
| temperature | 1 |
| top_p | 0.95 |
| concurrency | 8 |
AI-assisted source reading and original editorial synthesis; no independent model execution.