THE SOURCE IN CONTEXT
The selected seven-token draft arm is faster than the baseline in this source, but every arm retains quality_passed=false. The fastest three-token diagnostic failed the response contract before code execution. This is a text-only community experiment with useful negative results.
What is actually deployed?
The target and DFlash2 draft have separate full weight revisions, and the community image is pinned by digest. NVFP4 storage uses Marlin W4A16 MoE compute on GB10. The text-only TP2/RoCE setup is eager, uses FP8 E4M3 KV with an explicit 6 GiB allocation, a 131072 window and one request slot. It is not an NVIDIA-published GB10 production recipe, and the target export has no MTP head.
Source: Weschera · pinned document Weschera · pinned document
What workload defines the speed values?
Each arm has four naturally stopping requests: two English technical tutorials and two Korean essays. Input lengths are 78 or 92 tokens and generated lengths vary greatly. Median estimated decode excludes measured TTFT. One boot per arm and different output texts prevent a matched task-time or uncertainty claim. Saved request controls include temperature zero, top-p one and seed 37.
Source: Weschera · pinned document Weschera · pinned document
Why keep false quality flags?
The depth-7 arm passes the executed semantic checks but still returns a fenced block where bare Python was required and explanation where exactly 703 was required. The baseline and depth-3 arm fail the early one-code-block contract; later code checks are skipped rather than proven algorithmically wrong. All quality_passed flags remain false.
Source: Weschera · pinned document Weschera · pinned document
What do context and thinking checks establish?
All arms retrieve three beginning/middle/end markers from 120124 actual input tokens. That is smaller than the configured window and is not broad long-document reasoning. The custom forced-OFF template keeps maximum-effort system wording and leaks planning into final text, so native OFF is unqualified. The final restart has a health probe, not a repeated quality suite; whole-model peak RAM is unreported.
Source: Weschera · pinned document Weschera · 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.
| Configuration / workload | Metric | Reported value | Evidence |
|---|---|---|---|
| baseline median decode | decode rate | 15.181243819412654 tokens/smedian of four different prompts in one arm boot; not repeated identical task | Original row ↗ |
| dflash2-7 median decode | decode rate | 20.50006471628756 tokens/smedian of four different prompts in one arm boot; not repeated identical task | Original row ↗ |
| dflash2-3 median decode | decode rate | 24.77753188791761 tokens/smedian of four different prompts in one arm boot; not repeated identical task | Original row ↗ |
baseline median decode · decode rate · conditions
| model | GLM-5.3-Flash |
|---|---|
| model_artifact | nvidia/GLM-5.3-Flash-NVFP4 |
| model_revision | 423acf37583782c51c142d145aef733d72943d93 |
| runtime_execution_revision | Not specified |
| executed_image_digest | sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6 |
| runtime_patch_source | 050081dc41ce6edd4d3f15fa19dc3410ba4210e3 |
| engine | vLLM community GB10 image with Marlin W4A16 MoE |
| system | two 128 GB DGX Spark GB10, TP2 over RoCE; text only |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | NVIDIA NVFP4 storage; Marlin W4A16 MoE compute |
| kv_dtype | fp8_e4m3 |
| KV_allocation_GiB | 6 |
| context_limit_tokens | 131072 |
| configured_max_requests | 1 |
| concurrency | 1 |
| prefill_batch_tokens | 128 |
| enforce_eager | True |
| cache_state | prefix caching disabled |
| thinking | custom forced-OFF template; does not qualify as native OFF |
| temperature | 0 |
| top_p | 1 |
| seed | 37 |
| repetitions | four different sustained prompts in one boot per arm |
| experiment_date | Not specified |
| metric_denominator | completion tokens / (wall minus measured TTFT); estimated decode |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| peak_runtime_RAM | Not specified |
| draft_tokens | 0 |
| drafter | Not specified |
| drafter_revision | Not specified |
| input_tokens | [78, 92] |
| actual_completion_tokens_by_prompt | [4356, 2355, 10853, 2238] |
| all_speed_requests_natural_stop | True |
| quality_passed | False |
dflash2-7 median decode · decode rate · conditions
| model | GLM-5.3-Flash |
|---|---|
| model_artifact | nvidia/GLM-5.3-Flash-NVFP4 |
| model_revision | 423acf37583782c51c142d145aef733d72943d93 |
| runtime_execution_revision | Not specified |
| executed_image_digest | sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6 |
| runtime_patch_source | 050081dc41ce6edd4d3f15fa19dc3410ba4210e3 |
| engine | vLLM community GB10 image with Marlin W4A16 MoE |
| system | two 128 GB DGX Spark GB10, TP2 over RoCE; text only |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | NVIDIA NVFP4 storage; Marlin W4A16 MoE compute |
| kv_dtype | fp8_e4m3 |
| KV_allocation_GiB | 6 |
| context_limit_tokens | 131072 |
| configured_max_requests | 1 |
| concurrency | 1 |
| prefill_batch_tokens | 128 |
| enforce_eager | True |
| cache_state | prefix caching disabled |
| thinking | custom forced-OFF template; does not qualify as native OFF |
| temperature | 0 |
| top_p | 1 |
| seed | 37 |
| repetitions | four different sustained prompts in one boot per arm |
| experiment_date | Not specified |
| metric_denominator | completion tokens / (wall minus measured TTFT); estimated decode |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| peak_runtime_RAM | Not specified |
| draft_tokens | 7 |
| drafter | incoai/GLM-5.3-Flash-DFlash2 |
| drafter_revision | bf582e4eacc1810f76656d1811693ff6c6737d2a |
| input_tokens | [78, 92] |
| actual_completion_tokens_by_prompt | [7396, 2272, 2292, 2158] |
| all_speed_requests_natural_stop | True |
| quality_passed | False |
dflash2-3 median decode · decode rate · conditions
| model | GLM-5.3-Flash |
|---|---|
| model_artifact | nvidia/GLM-5.3-Flash-NVFP4 |
| model_revision | 423acf37583782c51c142d145aef733d72943d93 |
| runtime_execution_revision | Not specified |
| executed_image_digest | sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6 |
| runtime_patch_source | 050081dc41ce6edd4d3f15fa19dc3410ba4210e3 |
| engine | vLLM community GB10 image with Marlin W4A16 MoE |
| system | two 128 GB DGX Spark GB10, TP2 over RoCE; text only |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | NVIDIA NVFP4 storage; Marlin W4A16 MoE compute |
| kv_dtype | fp8_e4m3 |
| KV_allocation_GiB | 6 |
| context_limit_tokens | 131072 |
| configured_max_requests | 1 |
| concurrency | 1 |
| prefill_batch_tokens | 128 |
| enforce_eager | True |
| cache_state | prefix caching disabled |
| thinking | custom forced-OFF template; does not qualify as native OFF |
| temperature | 0 |
| top_p | 1 |
| seed | 37 |
| repetitions | four different sustained prompts in one boot per arm |
| experiment_date | Not specified |
| metric_denominator | completion tokens / (wall minus measured TTFT); estimated decode |
| output_tokens | Not specified |
| actual_output_tokens | Not specified |
| peak_runtime_RAM | Not specified |
| draft_tokens | 3 |
| drafter | incoai/GLM-5.3-Flash-DFlash2 |
| drafter_revision | bf582e4eacc1810f76656d1811693ff6c6737d2a |
| input_tokens | [78, 92] |
| actual_completion_tokens_by_prompt | [2822, 2194, 6331, 2105] |
| all_speed_requests_natural_stop | True |
| quality_passed | False |
Local Spark Qwen writing draft; Codex source audit and rewrite. No site model benchmark.