Qwen3.6 35B-A3B
Qwen/Qwen3.6-35B-A3B
Qwen · 9 recorded configurations
Measured setup
Hardware
Runtime and tool
Quantisation
MTP draft depth
Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.
1× RTX 3090 · llama.cpp
UD-Q3_K_XL · MTP off
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 3,227.23 | 138.87 |
| 2,048 | 2,663.26 | 135.78 |
| 4,096 | 2,602.61 | 133.83 |
| 8,192 | 2,533.83 | 129.05 |
| 16,384 | 2,399.47 | 119.65 |
| 32,768 | 2,097.01 | 105.33 |
| 65,536 | 1,733.91 | 84.16 |
| 131,072 | 1,285.76 | 60.23 |
| 261,312 | 833.70 | 38.11 |
Concurrent-stream measurements
| Streams | Context depth (tokens) | Aggregate PP (tok/s) | Aggregate TG (tok/s) | TG / stream (tok/s) | Aggregate all (tok/s) |
|---|---|---|---|---|---|
| 2 | 8,192 | 2,177.75 | 195.00 | 97.50 | 500.12 |
| 3 | 8,192 | 2,038.86 | 209.35 | 69.78 | 559.50 |
| 4 | 8,192 | 1,835.33 | 208.91 | 52.23 | 557.94 |
| 5 | 8,192 | 1,795.99 | 219.45 | 43.89 | 605.05 |
| 6 | 8,192 | 1,779.98 | 224.58 | 37.43 | 626.63 |
| 7 | 8,192 | 1,771.45 | 227.70 | 32.53 | 640.99 |
| 8 | 8,192 | 1,678.88 | 225.08 | 28.13 | 638.61 |
Technical record and provenance
Source 1 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-context-2026-10-14-single-gpu-isolated-off-c1
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- context
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 262,144 tokens
- Actual context per slot
- 262,144 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 0, 1 stream: PP mean 3,227.23 t/s (min 3,200.00, max 3,241.41, SD 19.26, n=3); TG mean 138.87 t/s (min 138.16, max 139.96, SD 0.78, n=3); All mean 377.26 t/s (min 375.92, max 379.55, SD 1.62, n=3) · KV 2,048, 1 stream: PP mean 2,663.26 t/s (min 2,656.47, max 2,672.19, SD 6.59, n=3); TG mean 135.78 t/s (min 135.44, max 136.42, SD 0.46, n=3); All mean 363.86 t/s (min 362.88, max 365.41, SD 1.11, n=3) · KV 4,096, 1 stream: PP mean 2,602.61 t/s (min 2,568.17, max 2,621.12, SD 24.38, n=3); TG mean 133.83 t/s (min 133.72, max 134.00, SD 0.12, n=3); All mean 358.28 t/s (min 357.68, max 358.86, SD 0.48, n=3) · KV 8,192, 1 stream: PP mean 2,533.83 t/s (min 2,521.84, max 2,540.48, SD 8.49, n=3); TG mean 129.05 t/s (min 128.66, max 129.39, SD 0.30, n=3); All mean 345.80 t/s (min 345.02, max 346.69, SD 0.69, n=3) · KV 16,384, 1 stream: PP mean 2,399.47 t/s (min 2,389.37, max 2,412.82, SD 9.84, n=3); TG mean 119.65 t/s (min 119.57, max 119.72, SD 0.06, n=3); All mean 321.32 t/s (min 321.26, max 321.37, SD 0.04, n=3) · KV 32,768, 1 stream: PP mean 2,097.01 t/s (min 2,067.38, max 2,138.39, SD 30.16, n=3); TG mean 105.33 t/s (min 105.11, max 105.60, SD 0.20, n=3); All mean 282.50 t/s (min 281.54, max 283.72, SD 0.91, n=3) · KV 65,536, 1 stream: PP mean 1,733.91 t/s (min 1,710.64, max 1,746.90, SD 16.49, n=3); TG mean 84.16 t/s (min 83.99, max 84.26, SD 0.12, n=3); All mean 226.53 t/s (min 225.89, max 226.97, SD 0.47, n=3) · KV 131,072, 1 stream: PP mean 1,285.76 t/s (min 1,284.55, max 1,287.52, SD 1.27, n=3); TG mean 60.23 t/s (min 60.17, max 60.31, SD 0.06, n=3); All mean 162.65 t/s (min 162.42, max 162.89, SD 0.19, n=3) · KV 261,312, 1 stream: PP mean 833.70 t/s (min 829.55, max 836.61, SD 3.01, n=3); TG mean 38.11 t/s (min 38.08, max 38.14, SD 0.02, n=3); All mean 103.20 t/s (min 103.14, max 103.26, SD 0.05, n=3)
- Measurement outcomes
- 0 tokens · 1 stream: measured; 2,048 tokens · 1 stream: measured; 4,096 tokens · 1 stream: measured; 8,192 tokens · 1 stream: measured; 16,384 tokens · 1 stream: measured; 32,768 tokens · 1 stream: measured; 65,536 tokens · 1 stream: measured; 131,072 tokens · 1 stream: measured; 261,312 tokens · 1 stream: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=1). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-14-approved-models-single-gpu/qwen36-q3-k-xl-context-isolated
Source 2 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c2
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 2 streams: PP mean 2,177.75 t/s (min 2,097.01, max 2,338.08, SD 113.37, n=3); TG mean 195.00 t/s (min 193.45, max 196.01, SD 1.11, n=3); All mean 500.12 t/s (min 494.01, max 507.73, SD 5.70, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=2). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 3 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c3
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 3 streams: PP mean 2,038.86 t/s (min 1,984.43, max 2,080.07, SD 40.15, n=3); TG mean 209.35 t/s (min 208.14, max 211.24, SD 1.35, n=3); All mean 559.50 t/s (min 555.77, max 562.46, SD 2.78, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=3). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 4 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c4
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 4 streams: PP mean 1,835.33 t/s (min 1,713.81, max 1,910.89, SD 86.77, n=3); TG mean 208.91 t/s (min 200.80, max 217.71, SD 6.92, n=3); All mean 557.94 t/s (min 546.75, max 565.10, SD 8.02, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=4). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 5 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c5
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 5 streams: PP mean 1,795.99 t/s (min 1,781.55, max 1,807.06, SD 10.68, n=3); TG mean 219.45 t/s (min 217.96, max 220.44, SD 1.07, n=3); All mean 605.05 t/s (min 601.29, max 607.49, SD 2.70, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=5). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 6 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c6
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 6 streams: PP mean 1,779.98 t/s (min 1,772.40, max 1,788.94, SD 6.82, n=3); TG mean 224.58 t/s (min 222.05, max 226.12, SD 1.80, n=3); All mean 626.63 t/s (min 619.95, max 630.63, SD 4.75, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=6). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 7 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c7
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 7 streams: PP mean 1,771.45 t/s (min 1,767.11, max 1,773.81, SD 3.08, n=3); TG mean 227.70 t/s (min 226.67, max 228.37, SD 0.74, n=3); All mean 640.99 t/s (min 638.35, max 642.74, SD 1.90, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=7). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 8 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-off-c8
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-off-02dba0cbd7
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 8 streams: PP mean 1,678.88 t/s (min 1,549.39, max 1,744.06, SD 91.56, n=3); TG mean 225.08 t/s (min 218.56, max 228.90, SD 4.63, n=3); All mean 638.61 t/s (min 620.92, max 648.94, SD 12.57, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, concurrency=8). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0