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 1
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 2,596.50 | 184.97 |
| 2,048 | 2,181.75 | 182.74 |
| 4,096 | 2,129.62 | 180.45 |
| 8,192 | 1,996.72 | 173.39 |
| 16,384 | 1,696.77 | 166.10 |
| 32,768 | 1,389.72 | 150.51 |
| 65,536 | 978.29 | 126.01 |
| 131,072 | 614.12 | 96.23 |
| 261,312 | 369.60 | 65.62 |
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 | 1,798.65 | 232.25 | 116.13 | 561.83 |
| 3 | 8,192 | 1,716.32 | 245.86 | 81.95 | 621.30 |
| 4 | 8,192 | 1,514.61 | 241.60 | 60.40 | 632.51 |
| 5 | 8,192 | 1,569.82 | 254.71 | 50.94 | 684.21 |
| 6 | 8,192 | 1,527.46 | 265.21 | 44.20 | 720.44 |
| 7 | 8,192 | 1,508.89 | 273.70 | 39.10 | 751.19 |
| 8 | 8,192 | 1,539.76 | 279.11 | 34.89 | 775.02 |
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-mtp-1-c1
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 2,596.50 t/s (min 2,559.69, max 2,621.46, SD 26.57, n=3); TG mean 184.97 t/s (min 183.95, max 185.96, SD 0.82, n=3); All mean 475.96 t/s (min 474.05, max 478.56, SD 1.90, n=3) · KV 2,048, 1 stream: PP mean 2,181.75 t/s (min 2,152.26, max 2,200.36, SD 21.09, n=3); TG mean 182.74 t/s (min 181.09, max 184.58, SD 1.43, n=3); All mean 460.22 t/s (min 456.99, max 464.57, SD 3.20, n=3) · KV 4,096, 1 stream: PP mean 2,129.62 t/s (min 2,112.93, max 2,143.14, SD 12.54, n=3); TG mean 180.45 t/s (min 180.18, max 180.71, SD 0.21, n=3); All mean 453.60 t/s (min 453.20, max 454.34, SD 0.52, n=3) · KV 8,192, 1 stream: PP mean 1,996.72 t/s (min 1,967.79, max 2,033.07, SD 27.16, n=3); TG mean 173.39 t/s (min 171.76, max 175.04, SD 1.34, n=3); All mean 434.41 t/s (min 430.83, max 438.86, SD 3.34, n=3) · KV 16,384, 1 stream: PP mean 1,696.77 t/s (min 1,644.51, max 1,726.42, SD 37.06, n=3); TG mean 166.10 t/s (min 165.19, max 166.57, SD 0.64, n=3); All mean 408.38 t/s (min 404.49, max 410.43, SD 2.75, n=3) · KV 32,768, 1 stream: PP mean 1,389.72 t/s (min 1,380.85, max 1,401.71, SD 8.80, n=3); TG mean 150.51 t/s (min 150.18, max 150.91, SD 0.30, n=3); All mean 363.53 t/s (min 362.42, max 364.82, SD 0.99, n=3) · KV 65,536, 1 stream: PP mean 978.29 t/s (min 933.23, max 1,006.09, SD 32.15, n=3); TG mean 126.01 t/s (min 125.41, max 126.69, SD 0.53, n=3); All mean 294.43 t/s (min 290.67, max 297.43, SD 2.81, n=3) · KV 131,072, 1 stream: PP mean 614.12 t/s (min 596.86, max 644.42, SD 21.50, n=3); TG mean 96.23 t/s (min 95.80, max 96.86, SD 0.46, n=3); All mean 215.31 t/s (min 213.21, max 218.88, SD 2.54, n=3) · KV 261,312, 1 stream: PP mean 369.60 t/s (min 358.28, max 375.89, SD 8.02, n=3); TG mean 65.62 t/s (min 65.38, max 65.80, SD 0.18, n=3); All mean 142.32 t/s (min 140.75, max 143.36, SD 1.13, 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 (mtp-1, 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-mtp-1-c2
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,798.65 t/s (min 1,797.66, max 1,800.21, SD 1.12, n=3); TG mean 232.25 t/s (min 231.97, max 232.46, SD 0.21, n=3); All mean 561.83 t/s (min 561.59, max 562.01, SD 0.18, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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-mtp-1-c3
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,716.32 t/s (min 1,652.80, max 1,773.36, SD 49.43, n=3); TG mean 245.86 t/s (min 245.15, max 246.49, SD 0.55, n=3); All mean 621.30 t/s (min 615.16, max 626.42, SD 4.65, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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-mtp-1-c4
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,514.61 t/s (min 1,441.03, max 1,565.56, SD 53.30, n=3); TG mean 241.60 t/s (min 238.36, max 243.45, SD 2.30, n=3); All mean 632.51 t/s (min 621.88, max 638.98, SD 7.57, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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-mtp-1-c5
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,569.82 t/s (min 1,550.94, max 1,596.18, SD 19.21, n=3); TG mean 254.71 t/s (min 253.94, max 255.22, SD 0.55, n=3); All mean 684.21 t/s (min 683.24, max 685.42, SD 0.91, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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-mtp-1-c6
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,527.46 t/s (min 1,526.21, max 1,529.64, SD 1.55, n=3); TG mean 265.21 t/s (min 264.02, max 266.01, SD 0.85, n=3); All mean 720.44 t/s (min 717.97, max 722.21, SD 1.80, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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-mtp-1-c7
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,508.89 t/s (min 1,487.65, max 1,524.07, SD 15.48, n=3); TG mean 273.70 t/s (min 272.68, max 275.45, SD 1.25, n=3); All mean 751.19 t/s (min 749.05, max 755.35, SD 2.95, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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-mtp-1-c8
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-1-272278ce62
- 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
- MTP · draft max 1
- 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 1,539.76 t/s (min 1,491.27, max 1,570.42, SD 34.68, n=3); TG mean 279.11 t/s (min 275.33, max 281.02, SD 2.68, n=3); All mean 775.02 t/s (min 762.04, max 782.70, SD 9.23, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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