Qwen3.8 27B
Qwen/Qwen3.8-27B
Qwen · 22 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 · ExLlamaV3
EXL3 5.00bpw · MTP 3
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 896.79 | 62.66 |
| 16,384 | 733.79 | 52.08 |
| 97,536 | 383.86 | 39.38 |
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 | 859.90 | 110.95 | 55.47 | 264.57 |
| 3 | 8,192 | 898.23 | 142.66 | 47.55 | 324.81 |
| 4 | 8,192 | 916.04 | 186.09 | 46.52 | 396.97 |
Technical record and provenance
Source 1 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-context-fit-1-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- context
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · context-fit 98304-token allocation
- Measured date
- 2026-10-15
- Hardware
- 1× RTX 3090
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- One physical RTX 3090; target model and native MTP compute tensors were observed on CUDA0, with no CPU expert offload or model split.
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Requested context per slot
- 98,304 tokens
- Actual context per slot
- 98,304 tokens
- Native model context
- 262,144 tokens
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Throughput spreads
- KV 0, 1 stream: PP mean 896.79 t/s (min 893.80, max 901.67, SD 4.26, n=3); TG mean 62.66 t/s (min 62.56, max 62.74, SD 0.09, n=3); All mean 164.93 t/s (min 164.66, max 165.09, SD 0.24, n=3) · KV 16,384, 1 stream: PP mean 733.79 t/s (min 733.00, max 735.01, SD 1.07, n=3); TG mean 52.08 t/s (min 52.00, max 52.17, SD 0.09, n=3); All mean 136.82 t/s (min 136.61, max 137.05, SD 0.22, n=3) · KV 97,536, 1 stream: PP mean 383.86 t/s (min 382.92, max 385.71, SD 1.61, n=3); TG mean 39.38 t/s (min 39.30, max 39.42, SD 0.07, n=3); All mean 98.02 t/s (min 97.82, max 98.17, SD 0.18, n=3)
- Measurement outcomes
- 0 tokens · 1 stream: measured; 16,384 tokens · 1 stream: measured; 97,536 tokens · 1 stream: measured
- Notes
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · context-fit 98304-token allocation. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. Runtime evidence maps that physical GPU to one logical CUDA0. Source configuration 500bpw-context-mtp-3-c1; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 6e4e9b07971a6b8fe0c0a54c2b4754b125a6c875aa696466ff060135cb56d69c. Separately allocated 98304 tokens per stream for this MTP setting. The allocation search advances in 8192-token steps; its recorded fitted capacity is not a token-exact VRAM maximum. The selected boundary and the 0/16K samples are retained from their recorded source attempts. The 106496-token allocation probe fit but its populated boundary failed; this export uses the independently verified 98304-token profile. The 114688-token allocation-only failure is not a populated-prefix boundary result.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-exl-context-fit/5bpw-gpu1/runs/mtp-3-98304-boundary/5.00bpw/context/mtp-3-c1
Source 2 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-2-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 2, 8192-token prefix, 8960-token allocation per stream
- Measured date
- 2026-10-15
- Hardware
- 1× RTX 3090
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- One physical RTX 3090; target model and native MTP compute tensors were observed on CUDA0, with no CPU expert offload or model split.
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Requested context per slot
- 8,960 tokens
- Actual context per slot
- 8,960 tokens
- Native model context
- 262,144 tokens
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Throughput spreads
- KV 8,192, 2 streams: PP mean 859.90 t/s (min 854.23, max 864.35, SD 5.17, n=3); TG mean 110.95 t/s (min 110.24, max 111.38, SD 0.62, n=3); All mean 264.57 t/s (min 262.88, max 265.67, SD 1.49, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 2, 8192-token prefix, 8960-token allocation per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. Runtime evidence maps that physical GPU to one logical CUDA0. Source configuration 500bpw-concurrency-mtp-3-c2; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation; this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c2
Source 3 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-3-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 3, 8192-token prefix, 8960-token allocation per stream
- Measured date
- 2026-10-15
- Hardware
- 1× RTX 3090
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- One physical RTX 3090; target model and native MTP compute tensors were observed on CUDA0, with no CPU expert offload or model split.
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Requested context per slot
- 8,960 tokens
- Actual context per slot
- 8,960 tokens
- Native model context
- 262,144 tokens
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Throughput spreads
- KV 8,192, 3 streams: PP mean 898.23 t/s (min 894.57, max 900.63, SD 3.22, n=3); TG mean 142.66 t/s (min 142.09, max 143.26, SD 0.59, n=3); All mean 324.81 t/s (min 323.50, max 326.05, SD 1.28, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 3, 8192-token prefix, 8960-token allocation per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. Runtime evidence maps that physical GPU to one logical CUDA0. Source configuration 500bpw-concurrency-mtp-3-c3; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation; this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c3
Source 4 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-4-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 4, 8192-token prefix, 8960-token allocation per stream
- Measured date
- 2026-10-15
- Hardware
- 1× RTX 3090
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- One physical RTX 3090; target model and native MTP compute tensors were observed on CUDA0, with no CPU expert offload or model split.
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Requested context per slot
- 8,960 tokens
- Actual context per slot
- 8,960 tokens
- Native model context
- 262,144 tokens
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Throughput spreads
- KV 8,192, 4 streams: PP mean 916.04 t/s (min 908.85, max 920.83, SD 6.34, n=3); TG mean 186.09 t/s (min 185.28, max 186.64, SD 0.72, n=3); All mean 396.97 t/s (min 394.85, max 398.41, SD 1.88, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 4, 8192-token prefix, 8960-token allocation per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. Runtime evidence maps that physical GPU to one logical CUDA0. Source configuration 500bpw-concurrency-mtp-3-c4; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation; this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c4
Source 5 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-5-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 5, 8192-token prefix, 8960-token allocation per stream
- Measured date
- 2026-10-15
- Hardware
- 1× RTX 3090
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- One physical RTX 3090; target model and native MTP compute tensors were observed on CUDA0, with no CPU expert offload or model split.
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Requested context per slot
- 8,960 tokens
- Actual context per slot
- 8,960 tokens
- Native model context
- 262,144 tokens
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Measurement outcomes
- 8,192 tokens · 5 streams: oom — OutOfMemoryError: CUDA out of memory. Tried to allocate 120.00 MiB. GPU 0 has a total capacity of 23.56 GiB of which 99.06 MiB is free. Including non-PyTorch memory, this process has 23.46 GiB memory in use. 23.56 GiB allowed; Of the allocated memory 23.09 GiB is allocated by PyTorch, and 41.95 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
- Notes
- 1× RTX3090 — no splitting · EXL3 5.00bpw · MTP-3 · concurrency 5, 8192-token prefix, 8960-token allocation per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. Runtime evidence maps that physical GPU to one logical CUDA0. Source configuration 500bpw-concurrency-mtp-3-c5; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation; this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c5
Source 6 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-6-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- Single RTX3090 requested — placement unverified · EXL3 5.00bpw · MTP-3 · concurrency 6, 8192-token prefix, requested 8960-token allocation (actual unknown) per stream
- Measured date
- 2026-10-15
- Hardware
- Not observed (outcome-only startup failure)
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- Unverified (startup failed)
- Observed devices
- None observed
- Requested devices
- CUDA0
- Split mode
- None
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Measurement outcomes
- 8,192 tokens · 6 streams: oom — RuntimeError: Insufficient VRAM in split for model and cache
- Notes
- Single RTX3090 requested — placement unverified · EXL3 5.00bpw · MTP-3 · concurrency 6, 8192-token prefix, requested 8960-token allocation (actual unknown) per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. No successful model placement was observed for this outcome-only source. Source configuration 500bpw-concurrency-mtp-3-c6; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation (actual allocation unknown); this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c6
Source 7 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-7-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- Single RTX3090 requested — placement unverified · EXL3 5.00bpw · MTP-3 · concurrency 7, 8192-token prefix, requested 8960-token allocation (actual unknown) per stream
- Measured date
- 2026-10-15
- Hardware
- Not observed (outcome-only startup failure)
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- Unverified (startup failed)
- Observed devices
- None observed
- Requested devices
- CUDA0
- Split mode
- None
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Measurement outcomes
- 8,192 tokens · 7 streams: oom — RuntimeError: Insufficient VRAM in split for model and cache
- Notes
- Single RTX3090 requested — placement unverified · EXL3 5.00bpw · MTP-3 · concurrency 7, 8192-token prefix, requested 8960-token allocation (actual unknown) per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. No successful model placement was observed for this outcome-only source. Source configuration 500bpw-concurrency-mtp-3-c7; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation (actual allocation unknown); this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c7
Source 8 · 2026-10-15
- Source run ID
- trpro-exl-qwen38-27b-5bpw-gpu1-5bpw-mtp-3-concurrency-8-20261015
- Page entry ID
- trpro-exl-qwen38-27b-5bpw-gpu1-mtp-3
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.8 27B EXL3 5.00bpw · MTP-3
- Configuration label
- Single RTX3090 requested — placement unverified · EXL3 5.00bpw · MTP-3 · concurrency 8, 8192-token prefix, requested 8960-token allocation (actual unknown) per stream
- Measured date
- 2026-10-15
- Hardware
- Not observed (outcome-only startup failure)
- Runtime
- exllamav3 · ExLlamaV3 1.5.0+cu128.torch2.9.0; native extension SHA-256 c1d02b410ef729618077ba287f01f05b229f31d118a7aae04a1b7e56142e3b2e
- Weight placement
- Unverified (startup failed)
- Observed devices
- None observed
- Requested devices
- CUDA0
- Split mode
- None
- KV-cache precision
- EXL native K8/V8 with QSA side planes and recurrent state
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- ExLlamaV3 persistent worker time.perf_counter_ns with torch.cuda.synchronize at phase boundaries; fresh PP excludes prefix preparation; aggregate TG is across streams
- Measurement outcomes
- 8,192 tokens · 8 streams: oom — RuntimeError: Insufficient VRAM in split for model and cache
- Notes
- Single RTX3090 requested — placement unverified · EXL3 5.00bpw · MTP-3 · concurrency 8, 8192-token prefix, requested 8960-token allocation (actual unknown) per stream. Persistent EXL native worker: 512 fresh prompt tokens and 256 generated tokens; one warmup then three measured repetitions. PP is fresh-prompt processing after the populated prefix; TG is aggregate decode across concurrent streams. Throughput means and sample standard deviations are recomputed from the retained raw repetitions. EXL native K8/V8 cache storage includes QSA side planes and recurrent state; this is not GGUF Q8_0. Physical GPU 1 (GPU-0a2bed40-39e4-e67f-5234-868704f9704b) was selected as the single device. No successful model placement was observed for this outcome-only source. Source configuration 500bpw-concurrency-mtp-3-c8; model revision a35e75a73baee51da709329d19294245cbeeb5d8; source manifest SHA-256 f889b720ea21d731842579e93d41eb5e4ed995a63475a7e69874a2330b529192. The concurrency source used an 8192-token prefix and a fixed 8960-token per-stream cache allocation (actual allocation unknown); this allocation is independent of context-fit capacities.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/trpro-exl-qwen38-27b-5bpw-gpu1-independent-gpus-validation-20261015/5.00bpw/concurrency/mtp-3-c8