THE SOURCE IN CONTEXT
One 128 GB unified-memory DGX Spark serves senfu NVFP4 at its full recorded revision through SGLang commit 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b. The recipe keeps PLE on NVMe through mmap, uses an FP8 dense path and an NVFP4 LM head, and disables MTP. The repository documentation pin is separate from the engine and weight revisions.
Which checkpoint and runtime were tested?
One 128 GB unified-memory DGX Spark serves senfu NVFP4 at its full recorded revision through SGLang commit 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b. The recipe keeps PLE on NVMe through mmap, uses an FP8 dense path and an NVFP4 LM head, and disables MTP. The repository documentation pin is separate from the engine and weight revisions.
Source: nabe2030 · pinned document
What workload defines the concurrency curve?
The selected ShareGPT-200 sweep uses vllm bench serve, thinking off and explicitly disabled radix caching, after one discarded warmup request. C1 is the median of three trials; C8, C16 and C32 are single trials. Exact prompt/output distributions, sampling and per-sweep window are not reported. All selected senfu trials completed 200 requests without errors.
Source: nabe2030 · pinned document
How should throughput and latency be read together?
Aggregate output rises with concurrency, while the reported median TTFT and time per output token also rise. Read all three metrics with their own timing definitions. Dividing aggregate rate by concurrency does not reconstruct a measured individual stream. Single high-load trials provide no variability estimate, and results from another checkpoint or engine cannot isolate a pure scheduler effect.
Source: nabe2030 · pinned document
Does the recipe prove simultaneous long windows?
The 49152-token window, 393216 total tokens and eight slots describe one published long-context configuration, not a saved all-slots-full workload. The separate 3.04 GB KV allocation at 128K keeps the original GB label; it is not whole-model peak RAM. The source reports 128K stability but supplies no complete long-request latency or memory series for every concurrency level.
Source: nabe2030 · pinned document
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.
single Spark senfu C1 aggregate_output_rate · aggregate output rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 1 |
| repetitions | 3 |
| summary_kind | median of three trials |
single Spark senfu C1 median_ttft · median ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 1 |
| repetitions | 3 |
| summary_kind | median of three trials |
single Spark senfu C1 median_time_per_output_token · median time per output token · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 1 |
| repetitions | 3 |
| summary_kind | median of three trials |
single Spark senfu C8 aggregate_output_rate · aggregate output rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 8 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C8 median_ttft · median ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 8 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C8 median_time_per_output_token · median time per output token · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 8 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C16 aggregate_output_rate · aggregate output rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 16 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C16 median_ttft · median ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 16 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C16 median_time_per_output_token · median time per output token · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 16 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C32 aggregate_output_rate · aggregate output rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 32 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C32 median_ttft · median ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 32 |
| repetitions | 1 |
| summary_kind | single trial |
single Spark senfu C32 median_time_per_output_token · median time per output token · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | senfu/Qwen3.8-Flash-Next-NVFP4 |
| model_revision | 5d37b3b3711d8406174b96ff950c0aa16324b266 |
| engine | SGLang |
| runtime_execution_revision | 4fb9b5b5bad41ac7de5586ec812c421fe3aa2b8b |
| executed_image_digest | Not specified |
| system | one DGX Spark GB10, 128 GB unified memory; NVMe file-backed PLE |
| node_count | 1 |
| weight_precision | NVFP4 experts and LM head; FP8 dense path |
| kv_dtype | Not specified |
| draft_tokens | 0 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| request_count | 200 |
| cache_state | radix/prefix cache explicitly disabled; one warmup discarded |
| harness | vllm bench serve, ShareGPT 200 prompts |
| experiment_date | Not specified |
| context_limit_tokens | Not specified |
| executed_context_for_each_sweep_unreported | True |
| PLE_backing | NVMe mmap |
| exact_prompt_and_output_distributions_unreported | True |
| concurrency | 32 |
| repetitions | 1 |
| summary_kind | single trial |
Original source-based interpretation; local Spark Qwen draft with Codex factual review, no site model benchmark.