THE SOURCE IN CONTEXT
The four selected C1 decode observations come from the raw quick-matrix CSVs, with 520-token thinking-off or 546-token effort-75 prompts and 64 actual output tokens. Each is one pass without variance estimates. Their first-to-last streamed-delta rate excludes initial prefill; end-to-end wall throughput remains a different field. This is not a long-context or sustained-response experiment.
What was measured?
The four selected C1 decode observations come from the raw quick-matrix CSVs, with 520-token thinking-off or 546-token effort-75 prompts and 64 actual output tokens. Each is one pass without variance estimates. Their first-to-last streamed-delta rate excludes initial prefill; end-to-end wall throughput remains a different field. This is not a long-context or sustained-response experiment.
Source: x00byte · pinned document x00byte · pinned document x00byte · pinned document
Which files and memory controls belong to this profile?
Both runtimes use the same pinned EXL3 conversion and separately pinned official 1017 Engrams. The TensorFold Python profile has eight request slots, a shared 524288-token pool, a 262144 per-request ceiling and configured 5/4 GiB memory floors. Floors are admission controls, not observed peak memory or a proof of full-window concurrency.
Source: x00byte · pinned document x00byte · pinned document x00byte · pinned document
Can these rates be used for a VQ or million-token deployment?
No such transfer is established. The runtime, quantized tensors, request length and cache policy must stay attached to each observation. The text matrix predates the explicit native-image-routing flag; a later red-image smoke check confirms that path only. No comprehensive image or model-quality benchmark is claimed.
How does storage placement affect the deployment description?
The reported worker reads checkpoint files over a read-only NFS share, while active Engram lookups remain rank-local. The storage inventory includes separate source tables and packed files. Disk preparation, model residency and a shared KV budget must be recorded independently rather than adding all on-disk bytes into a claimed live memory requirement.
Source: x00byte · 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 |
|---|---|---|---|
| tensorfold C1 thinking off | decode rate | 58.05409336065285 tokens/sone pass; no run-to-run uncertainty estimate | Original row ↗ |
| tensorfold C1 thinking 75 | decode rate | 77.49652509362433 tokens/sone pass; no run-to-run uncertainty estimate | Original row ↗ |
| vllm C1 thinking off | decode rate | 36.8753802051545 tokens/sone pass; no run-to-run uncertainty estimate | Original row ↗ |
| vllm C1 thinking 75 | decode rate | 41.447290817386964 tokens/sone pass; no run-to-run uncertainty estimate | Original row ↗ |
tensorfold C1 thinking off · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| Engram_shard_SHA256 | {"model-00047-of-00048.safetensors": "76a05106ece259be83d1bfea0dcb54f23377864ee60fd321690f35e6698a2af5", "model-00048-of-00048.safetensors": "9dc51433e195bc78b5ad846b68643cc0351d3b174e5e4309222d3b7402c83afd"} |
| weight_precision | EXL3 2.9bpw; 6-bit dense MTP handling in vLLM overlay |
| system | two DGX Spark GB10, 128 GB unified memory each, TP2, both available RoCE links |
| node_count | 2 |
| parallelism | TP2 |
| experiment_date | Not specified |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_depth_cap | 3 |
| concurrency | 1 |
| input_lines | 27 |
| output_tokens | 64 |
| max_output_tokens | 64 |
| repetitions | 1 |
| cache_state | startup and earlier generation checks; no per-case warmup; TensorFold session/prepared caches disabled |
| temperature | Not specified |
| top_p | Not specified |
| kv_dtype | Not specified |
| measurement_denominator | (completion_tokens - 1) / first-to-last streamed delta interval; excludes initial prefill |
| quality_evaluation | False |
| long_response_steady_state_measured | False |
| native_image_routing_in_text_matrix | False |
| engine | TensorFold Python 0.6.0 patched |
| runtime_execution_revision | c4646171139ee8a3c38103eaa1699dad226ec12b |
| overlay_source_revision | 2dcce0303451da2a59f80e6c7413da8b3b0f7a57 |
| executed_image_digest | Not specified |
| thinking | False |
| reasoning_effort | Not specified |
| input_tokens | 520 |
| actual_output_tokens | 64 |
| batch_wall_seconds | 2.4011132870218717 |
| end_to_end_output_wall_tps | 26.65430254620761 |
| TTFT_seconds | 1.313945789996069 |
tensorfold C1 thinking 75 · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| Engram_shard_SHA256 | {"model-00047-of-00048.safetensors": "76a05106ece259be83d1bfea0dcb54f23377864ee60fd321690f35e6698a2af5", "model-00048-of-00048.safetensors": "9dc51433e195bc78b5ad846b68643cc0351d3b174e5e4309222d3b7402c83afd"} |
| weight_precision | EXL3 2.9bpw; 6-bit dense MTP handling in vLLM overlay |
| system | two DGX Spark GB10, 128 GB unified memory each, TP2, both available RoCE links |
| node_count | 2 |
| parallelism | TP2 |
| experiment_date | Not specified |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_depth_cap | 3 |
| concurrency | 1 |
| input_lines | 27 |
| output_tokens | 64 |
| max_output_tokens | 64 |
| repetitions | 1 |
| cache_state | startup and earlier generation checks; no per-case warmup; TensorFold session/prepared caches disabled |
| temperature | Not specified |
| top_p | Not specified |
| kv_dtype | Not specified |
| measurement_denominator | (completion_tokens - 1) / first-to-last streamed delta interval; excludes initial prefill |
| quality_evaluation | False |
| long_response_steady_state_measured | False |
| native_image_routing_in_text_matrix | False |
| engine | TensorFold Python 0.6.0 patched |
| runtime_execution_revision | c4646171139ee8a3c38103eaa1699dad226ec12b |
| overlay_source_revision | 2dcce0303451da2a59f80e6c7413da8b3b0f7a57 |
| executed_image_digest | Not specified |
| thinking | True |
| reasoning_effort | 75 |
| input_tokens | 546 |
| actual_output_tokens | 64 |
| batch_wall_seconds | 2.141774678020738 |
| end_to_end_output_wall_tps | 29.881761446140466 |
| TTFT_seconds | 1.326543439005036 |
vllm C1 thinking off · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| Engram_shard_SHA256 | {"model-00047-of-00048.safetensors": "76a05106ece259be83d1bfea0dcb54f23377864ee60fd321690f35e6698a2af5", "model-00048-of-00048.safetensors": "9dc51433e195bc78b5ad846b68643cc0351d3b174e5e4309222d3b7402c83afd"} |
| weight_precision | EXL3 2.9bpw; 6-bit dense MTP handling in vLLM overlay |
| system | two DGX Spark GB10, 128 GB unified memory each, TP2, both available RoCE links |
| node_count | 2 |
| parallelism | TP2 |
| experiment_date | Not specified |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_depth_cap | 3 |
| concurrency | 1 |
| input_lines | 27 |
| output_tokens | 64 |
| max_output_tokens | 64 |
| repetitions | 1 |
| cache_state | startup and earlier generation checks; no per-case warmup; TensorFold session/prepared caches disabled |
| temperature | Not specified |
| top_p | Not specified |
| kv_dtype | Not specified |
| measurement_denominator | (completion_tokens - 1) / first-to-last streamed delta interval; excludes initial prefill |
| quality_evaluation | False |
| long_response_steady_state_measured | False |
| native_image_routing_in_text_matrix | False |
| engine | vLLM + EXL3 overlay |
| runtime_execution_revision | Not specified |
| overlay_source_revision | 6f7d1590ad49a2b8995188e45d7b9db31e677452 |
| executed_image_digest | sha256:2f0cf3adc0f989c1d446be274df864eb799630175f604c3b22b71b7205971dce |
| thinking | False |
| reasoning_effort | Not specified |
| input_tokens | 520 |
| actual_output_tokens | 64 |
| batch_wall_seconds | 2.5652991679962724 |
| end_to_end_output_wall_tps | 24.948357212460998 |
| TTFT_seconds | 0.8555283379973844 |
vllm C1 thinking 75 · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| Engram_shard_SHA256 | {"model-00047-of-00048.safetensors": "76a05106ece259be83d1bfea0dcb54f23377864ee60fd321690f35e6698a2af5", "model-00048-of-00048.safetensors": "9dc51433e195bc78b5ad846b68643cc0351d3b174e5e4309222d3b7402c83afd"} |
| weight_precision | EXL3 2.9bpw; 6-bit dense MTP handling in vLLM overlay |
| system | two DGX Spark GB10, 128 GB unified memory each, TP2, both available RoCE links |
| node_count | 2 |
| parallelism | TP2 |
| experiment_date | Not specified |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_depth_cap | 3 |
| concurrency | 1 |
| input_lines | 27 |
| output_tokens | 64 |
| max_output_tokens | 64 |
| repetitions | 1 |
| cache_state | startup and earlier generation checks; no per-case warmup; TensorFold session/prepared caches disabled |
| temperature | Not specified |
| top_p | Not specified |
| kv_dtype | Not specified |
| measurement_denominator | (completion_tokens - 1) / first-to-last streamed delta interval; excludes initial prefill |
| quality_evaluation | False |
| long_response_steady_state_measured | False |
| native_image_routing_in_text_matrix | False |
| engine | vLLM + EXL3 overlay |
| runtime_execution_revision | Not specified |
| overlay_source_revision | 6f7d1590ad49a2b8995188e45d7b9db31e677452 |
| executed_image_digest | sha256:2f0cf3adc0f989c1d446be274df864eb799630175f604c3b22b71b7205971dce |
| thinking | True |
| reasoning_effort | 75 |
| input_tokens | 546 |
| actual_output_tokens | 64 |
| batch_wall_seconds | 2.4199117650277913 |
| end_to_end_output_wall_tps | 26.447245277666145 |
| TTFT_seconds | 0.8985815939959139 |
Original source-grounded editorial synthesis with local Spark Qwen assistance and Codex review; no site model benchmark.