Gemma 4 12B
google/gemma-4-12B-it
Google · 4 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
Q8_0 · MTP 3
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 1,933.90 | 35.96 |
| 2,048 | 1,184.26 | 34.67 |
| 4,096 | 1,155.85 | 46.56 |
| 8,192 | 1,105.33 | 46.25 |
| 16,384 | 1,067.28 | 59.20 |
| 32,768 | 931.73 | 59.26 |
| 65,536 | 770.29 | 49.62 |
| 131,072 | 574.55 | 22.93 |
| 261,312 | 375.97 | 16.78 |
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 | 1,006.74 | 70.53 | 35.26 | 187.76 |
| 3 | 8,192 | 928.10 | 98.64 | 32.88 | 259.47 |
| 4 | 8,192 | 921.98 | 102.01 | 25.50 | 278.82 |
| 5 | 8,192 | 874.37 | 90.91 | 18.18 | 252.88 |
| 6 | 8,192 | 879.33 | 101.89 | 16.98 | 285.34 |
| 7 | 8,192 | 865.81 | 120.59 | 17.23 | 337.07 |
| 8 | 8,192 | 858.01 | 98.82 | 12.35 | 281.42 |
Technical record and provenance
Source 1 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-context-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c1
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- context
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 1,933.90 t/s (min 1,917.35, max 1,947.50, SD 12.48, n=3); TG mean 35.96 t/s (min 35.94, max 35.98, SD 0.02, n=3); All mean 103.56 t/s (min 103.47, max 103.65, SD 0.07, n=3) · KV 2,048, 1 stream: PP mean 1,184.26 t/s (min 1,172.38, max 1,193.75, SD 8.89, n=3); TG mean 34.67 t/s (min 34.64, max 34.69, SD 0.02, n=3); All mean 97.85 t/s (min 97.73, max 97.92, SD 0.09, n=3) · KV 4,096, 1 stream: PP mean 1,155.85 t/s (min 1,115.98, max 1,177.45, SD 28.23, n=3); TG mean 46.56 t/s (min 46.48, max 46.66, SD 0.07, n=3); All mean 128.53 t/s (min 128.12, max 128.97, SD 0.35, n=3) · KV 8,192, 1 stream: PP mean 1,105.33 t/s (min 1,065.32, max 1,125.95, SD 28.29, n=3); TG mean 46.25 t/s (min 46.23, max 46.27, SD 0.01, n=3); All mean 127.29 t/s (min 126.91, max 127.52, SD 0.27, n=3) · KV 16,384, 1 stream: PP mean 1,067.28 t/s (min 1,065.22, max 1,069.24, SD 1.64, n=3); TG mean 59.20 t/s (min 59.08, max 59.36, SD 0.12, n=3); All mean 158.63 t/s (min 158.35, max 158.98, SD 0.26, n=3) · KV 32,768, 1 stream: PP mean 931.73 t/s (min 910.02, max 944.41, SD 15.42, n=3); TG mean 59.26 t/s (min 59.19, max 59.31, SD 0.05, n=3); All mean 156.33 t/s (min 155.78, max 156.61, SD 0.39, n=3) · KV 65,536, 1 stream: PP mean 770.29 t/s (min 752.20, max 784.20, SD 13.39, n=3); TG mean 49.62 t/s (min 49.60, max 49.64, SD 0.01, n=3); All mean 130.66 t/s (min 130.29, max 130.94, SD 0.28, n=3) · KV 131,072, 1 stream: PP mean 574.55 t/s (min 572.97, max 576.66, SD 1.55, n=3); TG mean 22.93 t/s (min 22.91, max 22.95, SD 0.02, n=3); All mean 63.32 t/s (min 63.28, max 63.37, SD 0.04, n=3) · KV 261,312, 1 stream: PP mean 375.97 t/s (min 371.02, max 379.06, SD 3.53, n=3); TG mean 16.78 t/s (min 16.73, max 16.82, SD 0.04, n=3); All mean 45.92 t/s (min 45.75, max 46.04, SD 0.12, 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-3, 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/2026-10-15-gpu0-overlap-validation/gemma12-context-gpu0-validation
Source 2 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c2
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 1,006.74 t/s (min 935.82, max 1,105.89, SD 72.24, n=3); TG mean 70.53 t/s (min 69.19, max 73.15, SD 1.86, n=3); All mean 187.76 t/s (min 183.06, max 196.28, SD 6.03, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0
Source 3 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c3
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 928.10 t/s (min 911.35, max 944.86, SD 13.68, n=3); TG mean 98.64 t/s (min 98.47, max 98.89, SD 0.18, n=3); All mean 259.47 t/s (min 258.67, max 260.07, SD 0.59, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0
Source 4 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c4
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 921.98 t/s (min 897.49, max 964.22, SD 29.99, n=3); TG mean 102.01 t/s (min 100.62, max 104.28, SD 1.62, n=3); All mean 278.82 t/s (min 273.52, max 288.52, SD 6.87, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0
Source 5 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c5
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 874.37 t/s (min 859.35, max 888.86, SD 12.05, n=3); TG mean 90.91 t/s (min 90.52, max 91.14, SD 0.28, n=3); All mean 252.88 t/s (min 251.31, max 253.90, SD 1.12, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0
Source 6 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c6
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 879.33 t/s (min 878.36, max 879.86, SD 0.69, n=3); TG mean 101.89 t/s (min 101.63, max 102.06, SD 0.19, n=3); All mean 285.34 t/s (min 284.70, max 285.81, SD 0.47, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0
Source 7 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c7
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 865.81 t/s (min 859.35, max 873.65, SD 5.92, n=3); TG mean 120.59 t/s (min 120.46, max 120.84, SD 0.17, n=3); All mean 337.07 t/s (min 335.96, max 338.08, SD 0.87, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0
Source 8 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma12-q8-0-concurrency-2026-10-14-single-gpu-isolated-co-schedule-validation-mtp-3-c8
- Page entry ID
- gemma-4-12b-q8-0-single-gpu-20261016-mtp-3-7daeeaae74
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 12B Q8_0 — 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 3
- 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 858.01 t/s (min 835.76, max 871.73, SD 15.88, n=3); TG mean 98.82 t/s (min 98.09, max 99.25, SD 0.51, n=3); All mean 281.42 t/s (min 279.13, max 282.61, SD 1.62, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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/gemma12-q8-0-concurrency-physical0