Add cold-start benchmark results
Browse files- README.md +133 -0
- cross_model_report.md +194 -0
- observations.csv +142 -0
- runs.json +0 -0
README.md
ADDED
|
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- other
|
| 5 |
+
tags:
|
| 6 |
+
- benchmark
|
| 7 |
+
- inference
|
| 8 |
+
- cold-start
|
| 9 |
+
- llm-serving
|
| 10 |
+
- mlops
|
| 11 |
+
- systems
|
| 12 |
+
size_categories:
|
| 13 |
+
- n<1K
|
| 14 |
+
configs:
|
| 15 |
+
- config_name: default
|
| 16 |
+
data_files: observations.csv
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
# LLM Container Cold-Start Benchmark
|
| 20 |
+
|
| 21 |
+
Measurements of how long it takes to bring a language model from cold storage to
|
| 22 |
+
a state where it can serve its first token, across **25 open-weight
|
| 23 |
+
models** spanning **17 architecture families** and
|
| 24 |
+
**100.9 GiB** of checkpoints, on a single NVIDIA T4.
|
| 25 |
+
|
| 26 |
+
Cold start is the latency a serverless or scale-to-zero inference platform pays
|
| 27 |
+
when it has no warm replica. It decomposes into weight transfer from storage,
|
| 28 |
+
deserialization into host memory, transfer to the accelerator, and engine
|
| 29 |
+
bring-up. This dataset measures each phase separately rather than reporting one
|
| 30 |
+
aggregate number, because the dominant term shifts with model size and setup --
|
| 31 |
+
and optimising the wrong term buys nothing.
|
| 32 |
+
|
| 33 |
+
## Headline finding: cold start is superlinear in checkpoint size
|
| 34 |
+
|
| 35 |
+
Fitting `load_ms = a x GiB^b` gives **b = 1.58 to 2.10** across conditions, where b = 1 would mean constant throughput. **Doubling the checkpoint more than doubles the load time**, so no single MiB/s figure describes the system -- effective throughput is a function of model size.
|
| 36 |
+
|
| 37 |
+
Concretely, throughput holds near **1535 MiB/s** below **5.18 GiB** and falls to **758 MiB/s** above it -- a **2.02x** collapse.
|
| 38 |
+
|
| 39 |
+
Throughput in the healthy regime: **1354 MiB/s (R2=0.868)**.
|
| 40 |
+
|
| 41 |
+
The most likely cause is **host memory, not the storage device**: the runtime
|
| 42 |
+
had roughly 12-13 GiB of usable system RAM, and the memory-mapped load path
|
| 43 |
+
depends on the OS page cache holding the checkpoint while tensors are
|
| 44 |
+
materialised. This is stated as a hypothesis with an obvious test attached
|
| 45 |
+
(re-run on a high-RAM instance and see whether the boundary moves), not as a
|
| 46 |
+
settled conclusion.
|
| 47 |
+
|
| 48 |
+
**Why it matters:** a cold-start budget extrapolated from small models will be
|
| 49 |
+
optimistic for large ones, and the gap widens with size. Below the boundary,
|
| 50 |
+
storage bandwidth is the lever; above it, more storage bandwidth does not help,
|
| 51 |
+
because the bottleneck has moved.
|
| 52 |
+
|
| 53 |
+
## Other results
|
| 54 |
+
|
| 55 |
+
- **safetensors vs legacy pickle:** median **1.71x** faster to
|
| 56 |
+
load over 20 models with identical weights. The format costs no
|
| 57 |
+
model quality and no hardware to change.
|
| 58 |
+
- **Storage tier:** a bandwidth-capped tier is roughly an order of magnitude
|
| 59 |
+
slower than local NVMe on the same checkpoints.
|
| 60 |
+
- **Engine bring-up** is a comparable cost to weight transfer at these sizes,
|
| 61 |
+
which is precisely the distinction that determines whether snapshot/restore
|
| 62 |
+
techniques would pay off for a given deployment.
|
| 63 |
+
|
| 64 |
+
## Files
|
| 65 |
+
|
| 66 |
+
| File | Contents |
|
| 67 |
+
|---|---|
|
| 68 |
+
| `observations.csv` | One row per (model, experiment, condition), joined to checkpoint size, parameter count, shard count and family. |
|
| 69 |
+
| `runs.json` | The raw merged ledger, unmodified, so all statistics can be recomputed from source. |
|
| 70 |
+
| `cross_model_report.md` | The generated analysis: scaling fits, per-model tables, projections, noise profile, limitations. |
|
| 71 |
+
|
| 72 |
+
## Columns
|
| 73 |
+
|
| 74 |
+
`model_key`, `repo_id`, `family`, `tier`, `params_b`, `checkpoint_gib`,
|
| 75 |
+
`n_shards`, `bytes_per_param`, `experiment`, `condition`, `p50_ms`, `p95_ms`,
|
| 76 |
+
`stdev_ms`, `rsd`, `n_trials`, `throughput_mib_s`, `gpu`, `device_class`,
|
| 77 |
+
`reliable`.
|
| 78 |
+
|
| 79 |
+
`reliable` is `False` where relative standard deviation is 30% or more. Those
|
| 80 |
+
rows are published rather than dropped -- a reader may reasonably pick a
|
| 81 |
+
different threshold -- but no conclusion above rests on them.
|
| 82 |
+
|
| 83 |
+
## Method
|
| 84 |
+
|
| 85 |
+
- Each condition is repeated, warm-up runs are discarded, and **p50/p95** are
|
| 86 |
+
reported rather than a mean, because tail latency is what a scale-to-zero SLA
|
| 87 |
+
is written against.
|
| 88 |
+
- The OS page cache is evicted between trials (`posix_fadvise(DONTNEED)` on an
|
| 89 |
+
unprivileged runtime), so reads are genuinely cold rather than served from RAM.
|
| 90 |
+
- Runs were distributed across several Colab sessions coordinated through a
|
| 91 |
+
shared Postgres ledger with atomic claims and epoch fencing, so no task was
|
| 92 |
+
ever executed twice. Duplicate execution would not merely waste time -- two
|
| 93 |
+
workers would contend for the same disk and corrupt the I/O measurement.
|
| 94 |
+
- Registry metadata (checkpoint sizes, shard counts, gating) was read from the
|
| 95 |
+
Hugging Face API rather than estimated.
|
| 96 |
+
|
| 97 |
+
## Limitations
|
| 98 |
+
|
| 99 |
+
- **Free-tier hardware.** A single consumer GPU with shared, contended host I/O.
|
| 100 |
+
The *ratios* and the *scaling behaviour* transfer; the absolute milliseconds
|
| 101 |
+
are not production numbers.
|
| 102 |
+
- **Noise.** Median RSD is around 15%, with a long tail concentrated in engine
|
| 103 |
+
bring-up where JIT compilation varies run to run. Small effects here are not
|
| 104 |
+
findings.
|
| 105 |
+
- **Not measured:** driver-level checkpoint/restore (proprietary GPU memory
|
| 106 |
+
snapshotting), multi-GPU tensor-parallel loading, and real network-attached
|
| 107 |
+
storage -- the remote tier is a software bandwidth cap, which bounds the
|
| 108 |
+
penalty from below rather than reproducing it.
|
| 109 |
+
- **Coverage gaps** are listed at the top of the report.
|
| 110 |
+
|
| 111 |
+
## Reproducing
|
| 112 |
+
- **Code:** https://github.com/priyansh-saxena1/coldstart-lab
|
| 113 |
+
|
| 114 |
+
```bash
|
| 115 |
+
pip install -e ".[distributed,plots]"
|
| 116 |
+
coldstart-fleet init --tier small --device-class t4
|
| 117 |
+
coldstart-fleet work --device cuda --device-class t4
|
| 118 |
+
coldstart-fleet merge --out ./merged
|
| 119 |
+
```
|
| 120 |
+
|
| 121 |
+
## Citation
|
| 122 |
+
|
| 123 |
+
```bibtex
|
| 124 |
+
@misc{coldstart_bench_2026,
|
| 125 |
+
title = {LLM Container Cold-Start Benchmark},
|
| 126 |
+
year = {2026},
|
| 127 |
+
note = {Dataset: https://huggingface.co/datasets/ArchCoder/llm-cold-start-benchmark}
|
| 128 |
+
}
|
| 129 |
+
```
|
| 130 |
+
|
| 131 |
+
## License
|
| 132 |
+
|
| 133 |
+
Apache-2.0. Measurements only; no model weights are redistributed.
|
cross_model_report.md
ADDED
|
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# LLM Container Cold-Start: Cross-Model Report
|
| 2 |
+
|
| 3 |
+
## 1. Coverage
|
| 4 |
+
|
| 5 |
+
- Models benchmarked: **25**
|
| 6 |
+
- Observations: **141**
|
| 7 |
+
- Weights moved: **100.9 GiB**
|
| 8 |
+
- Architecture families: 17 (bloom, deepseek, deepseek-r1, falcon3, gpt2, granite, phi, pythia...)
|
| 9 |
+
- Hardware: Tesla T4
|
| 10 |
+
- Experiments: checkpoint_format (70), engine_init (23), storage_tier (48)
|
| 11 |
+
|
| 12 |
+
**Coverage gaps (3 models incomplete).** Stated up front because a missing cell is a silent hole in every table below:
|
| 13 |
+
|
| 14 |
+
- `granite-3.1-2b`: missing storage_tier
|
| 15 |
+
- `qwen2.5-7b-awq`: missing engine_init
|
| 16 |
+
- `qwen2.5-7b-gptq-int4`: missing engine_init
|
| 17 |
+
|
| 18 |
+
## 2. How does load time scale with checkpoint size?
|
| 19 |
+
|
| 20 |
+
Three fits per condition, because the first one alone is misleading. `ms/GiB` is the familiar linear slope. `exponent b` comes from a log-log fit of `ms = a x GiB^b`: b~1 means constant throughput, b>1 means throughput *degrades* as checkpoints grow. `fixed cost` is the linear model's intercept -- a checkpoint of size zero cannot take negative time, so a large negative value is proof the linear model is the wrong shape.
|
| 21 |
+
|
| 22 |
+
| Condition | ms per GiB | exponent b | linear R2 | fixed cost (ms) | n |
|
| 23 |
+
|---|---:|---:|---:|---:|---:|
|
| 24 |
+
| safetensors-nommap | 2187 | 1.58 | 0.676 | -4237 :warning: | 25 |
|
| 25 |
+
| safetensors | 3271 | 1.79 | 0.663 | -7001 :warning: | 25 |
|
| 26 |
+
| pytorch_bin | 6225 | 2.10 | 0.846 | -12263 :warning: | 20 |
|
| 27 |
+
|
| 28 |
+
### 2.1 The headline finding: cold start is superlinear
|
| 29 |
+
|
| 30 |
+
Every condition has b > 1 (up to **b = 2.10** for `pytorch_bin`). Doubling the checkpoint more than doubles the load time, so there is no single MiB/s that describes this system -- effective throughput is a function of model size, not a constant of the hardware.
|
| 31 |
+
|
| 32 |
+
Every linear fixed cost above is sharply negative, which is not a small numerical artefact: it is the linear model bending to chase a curve. Quoting those slopes as throughput, or extrapolating from them, would be quoting an artefact.
|
| 33 |
+
|
| 34 |
+
### 2.2 Where throughput collapses
|
| 35 |
+
|
| 36 |
+
Taking the median throughput of the smallest checkpoints as a healthy baseline and walking upward until it falls below 75% of that:
|
| 37 |
+
|
| 38 |
+
| Condition | healthy MiB/s | collapses above | degraded MiB/s | drop |
|
| 39 |
+
|---|---:|---:|---:|---:|
|
| 40 |
+
| pytorch_bin | 1171 | 1.40 GiB | 886 | 1.3x |
|
| 41 |
+
| safetensors | 1535 | 5.18 GiB | 758 | 2.0x |
|
| 42 |
+
| safetensors-nommap | 1562 | 5.18 GiB | 841 | 1.9x |
|
| 43 |
+
|
| 44 |
+
**The most likely cause is host memory, not the storage device.** These runs used a Colab T4 instance with roughly 12-13 GiB of usable system RAM. The mmap path depends on the OS page cache holding the checkpoint while tensors are materialised; once a checkpoint plus its in-flight copies approaches that budget, pages are evicted and re-read and the effective rate falls off. The collapse point sitting near 5 GiB -- comfortably under total RAM, but not under RAM minus the working copy -- is consistent with that reading.
|
| 45 |
+
|
| 46 |
+
This is a testable claim rather than a conclusion, and the test is cheap: re-run the same models on a high-RAM instance. If the boundary moves with available memory it is memory pressure; if it stays at 5 GiB it is something in the loader.
|
| 47 |
+
|
| 48 |
+
**Why it matters for a serving platform:** a cold-start budget extrapolated from small models will be optimistic for large ones, and the gap widens with size. It also means the first question about any deployment is which regime it sits in -- below the boundary, storage bandwidth is the lever; above it, no amount of storage bandwidth helps because the bottleneck has moved to memory.
|
| 49 |
+
|
| 50 |
+
## 3. Checkpoint format: safetensors vs legacy pickle
|
| 51 |
+
|
| 52 |
+
| Model | GiB | safetensors (ms) | .bin (ms) | speedup |
|
| 53 |
+
|---|---:|---:|---:|---:|
|
| 54 |
+
| `bloom-1b7` | 3.208 | 2017.4 | 3475.4 | 1.72x |
|
| 55 |
+
| `bloomz-560m` | 1.042 | 695.2 | 902.5 | 1.3x |
|
| 56 |
+
| `deepseek-coder-1.3b` | 2.508 | 1715.6 | 2193.0 | 1.28x |
|
| 57 |
+
| `deepseek-r1-distill-qwen-1.5b` | 3.31 | 2872.8 | 3407.9 | 1.19x |
|
| 58 |
+
| `falcon3-1b` | 3.11 | 2086.5 | 3594.2 | 1.72x |
|
| 59 |
+
| `gpt2-large` | 3.024 | 2128.8 | 3598.9 | 1.69x |
|
| 60 |
+
| `gpt2-medium` | 1.416 | 890.6 | 1166.8 | 1.31x |
|
| 61 |
+
| `granite-3.1-2b` | 4.719 | 4069.4 | 21260.6 | 5.22x |
|
| 62 |
+
| `phi-2` | 5.178 | 4854.0 | 23573.9 | 4.86x |
|
| 63 |
+
| `pythia-1.4b` | 2.729 | 1723.5 | 3503.2 | 2.03x |
|
| 64 |
+
| `pythia-2.8b` | 5.294 | 6273.7 | 23810.8 | 3.8x |
|
| 65 |
+
| `qwen2.5-1.5b` | 2.875 | 1807.3 | 2783.4 | 1.54x |
|
| 66 |
+
| `qwen2.5-coder-1.5b` | 2.875 | 2423.9 | 2740.7 | 1.13x |
|
| 67 |
+
| `qwen3-0.6b` | 1.4 | 949.5 | 1850.1 | 1.95x |
|
| 68 |
+
| `qwen3-1.7b` | 3.784 | 2394.4 | 4360.8 | 1.82x |
|
| 69 |
+
| `smollm2-1.7b` | 3.188 | 1980.5 | 3392.7 | 1.71x |
|
| 70 |
+
| `smollm3-3b` | 5.728 | 7371.8 | 25872.5 | 3.51x |
|
| 71 |
+
| `stablelm-2-1-6b` | 6.126 | 8701.5 | 27647.7 | 3.18x |
|
| 72 |
+
| `stablelm-zephyr-3b` | 5.207 | 20583.5 | 21731.0 | 1.06x |
|
| 73 |
+
| `tinyllama-1.1b-v1.0` | 2.049 | 1315.0 | 1988.3 | 1.51x |
|
| 74 |
+
|
| 75 |
+
Median speedup **1.71x** across 20 models (range 1.06-5.22x). Same tensors, same bytes on the wire: the difference is that safetensors memory-maps a flat buffer while the pickle path reconstructs every tensor through the Python interpreter. This is the cheapest available win -- a format migration costs no model quality and no hardware.
|
| 76 |
+
|
| 77 |
+
## 4. Quantization: what 4-bit buys at *load* time
|
| 78 |
+
|
| 79 |
+
_Run a quantized/fp16 pair (e.g. `qwen2.5-7b` and `qwen2.5-7b-awq`) to populate this section._
|
| 80 |
+
|
| 81 |
+
## 5. Per-shard overhead
|
| 82 |
+
|
| 83 |
+
| Checkpoint layout | ms per GiB | fixed cost (ms) | R2 | n |
|
| 84 |
+
|---|---:|---:|---:|---:|
|
| 85 |
+
| single shard | 2971.4 | -5267.0 | 0.5201 | 15 |
|
| 86 |
+
| multi shard | 6688.9 | -28010.1 | 0.8451 | 10 |
|
| 87 |
+
|
| 88 |
+
Extra fixed cost attributable to sharding: **-22743.1 ms**. On a local filesystem this is just extra `open()`s and header parses. On network-attached or object storage each shard is a separate request with its own round-trip, so this term grows with latency and is worth re-measuring against the real backing store.
|
| 89 |
+
|
| 90 |
+
## 6. Storage tiers
|
| 91 |
+
|
| 92 |
+
| Tier | median MiB/s | min | max | models |
|
| 93 |
+
|---|---:|---:|---:|---:|
|
| 94 |
+
| local-nvme | 1332.6 | 241.3 | 1638.9 | 24 |
|
| 95 |
+
| remote-emulated-200MiBs | 172.5 | 109.5 | 178.3 | 24 |
|
| 96 |
+
|
| 97 |
+
Tiers whose name contains `emulated` are bandwidth-capped in software rather than measured against real remote storage; the harness only ever slows a read down, never speeds it up, so these are a floor on the real penalty.
|
| 98 |
+
|
| 99 |
+
## 7. Engine bring-up
|
| 100 |
+
|
| 101 |
+
| Model | GiB | backend | p50 (ms) | p95 (ms) |
|
| 102 |
+
|---|---:|---|---:|---:|
|
| 103 |
+
| `bloomz-560m` | 1.042 | transformers | 2870.1 | 3227.8 |
|
| 104 |
+
| `qwen3-0.6b` | 1.4 | transformers | 3109.5 | 7340.8 |
|
| 105 |
+
| `gpt2-medium` | 1.416 | transformers | 1285.0 | 5730.4 |
|
| 106 |
+
| `tinyllama-1.1b-v1.0` | 2.049 | transformers | 3825.7 | 6858.9 |
|
| 107 |
+
| `deepseek-coder-1.3b` | 2.508 | transformers | 1947.2 | 7655.4 |
|
| 108 |
+
| `pythia-1.4b` | 2.729 | transformers | 1334.5 | 3878.4 |
|
| 109 |
+
| `qwen2.5-1.5b` | 2.875 | transformers | 2262.2 | 5195.4 |
|
| 110 |
+
| `qwen2.5-coder-1.5b` | 2.875 | transformers | 4862.0 | 10771.0 |
|
| 111 |
+
| `gpt2-large` | 3.024 | transformers | 4213.0 | 10072.8 |
|
| 112 |
+
| `falcon3-1b` | 3.11 | transformers | 3882.9 | 12556.5 |
|
| 113 |
+
| `smollm2-1.7b` | 3.188 | transformers | 1520.0 | 7757.8 |
|
| 114 |
+
| `bloom-1b7` | 3.208 | transformers | 10647.4 | 10798.6 |
|
| 115 |
+
| `deepseek-r1-distill-qwen-1.5b` | 3.31 | transformers | 3525.1 | 13968.9 |
|
| 116 |
+
| `qwen3-1.7b` | 3.784 | transformers | 8692.3 | 18968.1 |
|
| 117 |
+
| `granite-3.1-2b` | 4.719 | transformers | 12946.3 | 19295.6 |
|
| 118 |
+
| `phi-2` | 5.178 | transformers | 17282.6 | 19939.7 |
|
| 119 |
+
| `stablelm-zephyr-3b` | 5.207 | transformers | 19660.7 | 20809.7 |
|
| 120 |
+
| `pythia-2.8b` | 5.294 | transformers | 19414.5 | 21810.3 |
|
| 121 |
+
| `smollm3-3b` | 5.728 | transformers | 22397.2 | 23205.7 |
|
| 122 |
+
| `stablelm-2-1-6b` | 6.126 | transformers | 19685.1 | 21764.4 |
|
| 123 |
+
| `phi-3.5-mini` | 7.117 | transformers | 25219.9 | 30470.2 |
|
| 124 |
+
| `phi-3-mini-4k` | 7.117 | transformers | 15268.9 | 22811.1 |
|
| 125 |
+
| `qwen3-4b` | 7.492 | transformers | 32412.5 | 32641.3 |
|
| 126 |
+
|
| 127 |
+
Weight transfer is only one term. Engine bring-up adds CUDA context creation, kernel/JIT warm-up, graph capture and KV-cache allocation. Where this term is large relative to weight loading, snapshot/restore techniques pay off; where weight transfer dominates, they do not -- which is exactly the distinction this table is here to make.
|
| 128 |
+
|
| 129 |
+
## 8. Projection to production scale
|
| 130 |
+
|
| 131 |
+
Projecting from the **stable regime only** (1354 MiB/s, R2=0.868, n=16, checkpoints below 5.18 GiB), fitted through the origin so a checkpoint of size zero costs zero time.
|
| 132 |
+
|
| 133 |
+
The degraded regime is deliberately excluded: past the boundary the slope describes this instance running out of host memory, and a production node has a different amount of memory. Projecting the collapse would be projecting a property of a free Colab VM onto a server.
|
| 134 |
+
|
| 135 |
+
| Model | GiB | params | predicted weight load |
|
| 136 |
+
|---|---:|---:|---:|
|
| 137 |
+
| `phi-3-medium-4k` | 26.003 | 14.0B | 20 s |
|
| 138 |
+
| `phi-4` | 27.305 | 14.7B | 21 s |
|
| 139 |
+
| `qwen3-14b` | 27.508 | 14.8B | 21 s |
|
| 140 |
+
| `qwen2.5-14b` | 27.511 | 14.8B | 21 s |
|
| 141 |
+
| `qwen3-30b-a3b` | 56.873 | 30.5B | 43 s |
|
| 142 |
+
| `qwen3-32b` | 61.024 | 32.8B | 46 s |
|
| 143 |
+
| `qwen2.5-32b` | 61.028 | 32.5B | 46 s |
|
| 144 |
+
| `qwen2.5-72b` | 135.426 | 72.7B | 102 s |
|
| 145 |
+
|
| 146 |
+
**These are a lower bound, and both directions of error are known:**
|
| 147 |
+
|
| 148 |
+
- *Optimistic*, because it assumes throughput stays healthy at sizes far beyond anything measured here. Whether it does depends on the target having enough memory headroom -- exactly the question section 2.2 raises.
|
| 149 |
+
- *Pessimistic*, because production serving shards a large checkpoint across several GPUs and pulls the shards in parallel, while every number here is a single-device sequential load.
|
| 150 |
+
- It covers **weight transfer only**. Section 7 shows engine bring-up is a comparable cost at these sizes; a real cold start is the sum.
|
| 151 |
+
|
| 152 |
+
For a 27-32B class model the arithmetic gives roughly 46 s of sequential weight transfer at this throughput. That is a sanity check on an order of magnitude, not a prediction for any particular deployment.
|
| 153 |
+
|
| 154 |
+
## 9. How much to trust these numbers
|
| 155 |
+
|
| 156 |
+
- Median relative standard deviation: **15.4%**
|
| 157 |
+
- 90th percentile RSD: 82.9%
|
| 158 |
+
- Worst case RSD: 215.4%
|
| 159 |
+
- Conditions measured: 141
|
| 160 |
+
|
| 161 |
+
An effect smaller than the run-to-run spread is not a finding. Read every speedup above against this table: differences of a few percent are noise, and only the large multiples are safe to act on. Cold reads were forced between trials (page-cache eviction, or `posix_fadvise` where unprivileged), and warm-up runs discarded.
|
| 162 |
+
|
| 163 |
+
### 9.1 Measurements to discount (41 of 141, 29%)
|
| 164 |
+
|
| 165 |
+
These conditions have a relative standard deviation of 30% or more. They are reported rather than silently dropped, but no conclusion above rests on them, and they are the first candidates for a re-run with more repeats.
|
| 166 |
+
|
| 167 |
+
| Model | Experiment | Condition | p50 (ms) | RSD |
|
| 168 |
+
|---|---|---|---:|---:|
|
| 169 |
+
| `smollm2-1.7b` | engine_init | transformers | 1520.0 | 215% |
|
| 170 |
+
| `gpt2-medium` | engine_init | transformers | 1285.0 | 187% |
|
| 171 |
+
| `deepseek-coder-1.3b` | engine_init | transformers | 1947.2 | 161% |
|
| 172 |
+
| `deepseek-r1-distill-qwen-1.5b` | engine_init | transformers | 3525.1 | 160% |
|
| 173 |
+
| `falcon3-1b` | engine_init | transformers | 3882.9 | 128% |
|
| 174 |
+
| `bloom-1b7` | checkpoint_format | safetensors | 2017.4 | 118% |
|
| 175 |
+
| `phi-2` | checkpoint_format | safetensors | 4854.0 | 112% |
|
| 176 |
+
| `deepseek-coder-1.3b` | checkpoint_format | safetensors | 1715.6 | 109% |
|
| 177 |
+
| `pythia-1.4b` | engine_init | transformers | 1334.5 | 101% |
|
| 178 |
+
| `qwen2.5-7b-awq` | storage_tier | local-nvme | 5635.8 | 94% |
|
| 179 |
+
| `gpt2-large` | engine_init | transformers | 4213.0 | 89% |
|
| 180 |
+
| `deepseek-r1-distill-qwen-1.5b` | checkpoint_format | safetensors | 2872.8 | 88% |
|
| 181 |
+
| `qwen3-1.7b` | storage_tier | local-nvme | 2967.4 | 88% |
|
| 182 |
+
| `qwen2.5-7b-gptq-int4` | checkpoint_format | safetensors | 5751.2 | 86% |
|
| 183 |
+
| `qwen3-1.7b` | engine_init | transformers | 8692.3 | 83% |
|
| 184 |
+
| _... and 26 more_ | | | | |
|
| 185 |
+
|
| 186 |
+
The dominant source is almost certainly the shared, contended I/O of a free-tier VM: neighbouring tenants on the same host move the storage numbers far more than anything in the harness does. This is the strongest argument for treating the *ratios* here as the result and the absolute milliseconds as context.
|
| 187 |
+
|
| 188 |
+
## 10. Limitations
|
| 189 |
+
|
| 190 |
+
- Free-tier hardware: a single consumer GPU with shared, contended host I/O. Absolute numbers are not production numbers; the *ratios* and the *scaling behaviour* are what transfer.
|
| 191 |
+
- Driver-level checkpoint/restore (proprietary GPU memory snapshotting) cannot be reproduced here. Where a technique was out of reach, the closest honest proxy was measured and labelled.
|
| 192 |
+
- Emulated storage tiers are software bandwidth caps, not real network storage.
|
| 193 |
+
- Projections assume linear scaling in bytes; see section 8.
|
| 194 |
+
|
observations.csv
ADDED
|
@@ -0,0 +1,142 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
model_key,repo_id,family,tier,params_b,checkpoint_gib,n_shards,bytes_per_param,experiment,condition,p50_ms,p95_ms,stdev_ms,rsd,n_trials,throughput_mib_s,gpu,device_class,reliable
|
| 2 |
+
bloomz-560m,bigscience/bloomz-560m,bloom,small,0.56,1.042,1,1.998,checkpoint_format,pytorch_bin,902.46,1190.88,139.69,0.1548,5,1182.34,Tesla T4,t4,True
|
| 3 |
+
qwen3-0.6b,Qwen/Qwen3-0.6B,qwen3,small,0.6,1.4,1,2.505,checkpoint_format,pytorch_bin,1850.08,2485.44,422.23,0.2282,5,774.89,Tesla T4,t4,True
|
| 4 |
+
gpt2-medium,openai-community/gpt2-medium,gpt2,small,0.355,1.416,1,4.283,checkpoint_format,pytorch_bin,1166.8,1244.53,39.12,0.0335,5,1242.71,Tesla T4,t4,True
|
| 5 |
+
tinyllama-1.1b-v1.0,TinyLlama/TinyLlama-1.1B-Chat-v1.0,tinyllama,small,1.1,2.049,1,2.0,checkpoint_format,pytorch_bin,1988.33,2160.62,231.27,0.1163,5,1055.24,Tesla T4,t4,True
|
| 6 |
+
deepseek-coder-1.3b,deepseek-ai/deepseek-coder-1.3b-instruct,deepseek,small,1.35,2.508,1,1.995,checkpoint_format,pytorch_bin,2192.96,2620.15,215.68,0.0984,5,1171.11,Tesla T4,t4,True
|
| 7 |
+
pythia-1.4b,EleutherAI/pythia-1.4b,pythia,small,1.4,2.729,1,2.093,checkpoint_format,pytorch_bin,3503.16,4721.1,598.47,0.1708,5,797.71,Tesla T4,t4,True
|
| 8 |
+
qwen2.5-1.5b,Qwen/Qwen2.5-1.5B-Instruct,qwen2.5,small,1.54,2.875,1,2.005,checkpoint_format,pytorch_bin,2783.45,4950.61,1080.88,0.3883,5,1057.68,Tesla T4,t4,False
|
| 9 |
+
qwen2.5-coder-1.5b,Qwen/Qwen2.5-Coder-1.5B-Instruct,qwen2.5-coder,small,1.54,2.875,1,2.005,checkpoint_format,pytorch_bin,2740.7,3540.9,428.0,0.1562,5,1074.18,Tesla T4,t4,True
|
| 10 |
+
gpt2-large,openai-community/gpt2-large,gpt2,small,0.774,3.024,1,4.195,checkpoint_format,pytorch_bin,3598.93,5390.23,835.21,0.2321,5,860.42,Tesla T4,t4,True
|
| 11 |
+
falcon3-1b,tiiuae/Falcon3-1B-Instruct,falcon3,small,1.0,3.11,1,3.339,checkpoint_format,pytorch_bin,3594.21,6537.42,1394.48,0.388,5,886.05,Tesla T4,t4,False
|
| 12 |
+
smollm2-1.7b,HuggingFaceTB/SmolLM2-1.7B-Instruct,smollm2,small,1.71,3.188,1,2.002,checkpoint_format,pytorch_bin,3392.74,5465.0,980.37,0.289,5,962.21,Tesla T4,t4,True
|
| 13 |
+
bloom-1b7,bigscience/bloom-1b7,bloom,small,1.7,3.208,1,2.026,checkpoint_format,pytorch_bin,3475.35,7047.18,1786.19,0.514,5,945.23,Tesla T4,t4,False
|
| 14 |
+
deepseek-r1-distill-qwen-1.5b,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,deepseek-r1,small,1.78,3.31,1,1.997,checkpoint_format,pytorch_bin,3407.89,8064.1,2245.43,0.6589,5,994.59,Tesla T4,t4,False
|
| 15 |
+
qwen3-1.7b,Qwen/Qwen3-1.7B,qwen3,small,1.7,3.784,2,2.39,checkpoint_format,pytorch_bin,4360.76,8873.03,2432.73,0.5579,5,888.56,Tesla T4,t4,False
|
| 16 |
+
granite-3.1-2b,ibm-granite/granite-3.1-2b-instruct,granite,small,2.53,4.719,2,2.003,checkpoint_format,pytorch_bin,21260.61,21322.24,43.08,0.002,5,227.29,Tesla T4,t4,True
|
| 17 |
+
phi-2,microsoft/phi-2,phi,small,2.78,5.178,2,2.0,checkpoint_format,pytorch_bin,23573.88,23775.34,148.26,0.0063,5,224.92,Tesla T4,t4,True
|
| 18 |
+
stablelm-zephyr-3b,stabilityai/stablelm-zephyr-3b,stablelm,small,2.8,5.207,1,1.997,checkpoint_format,pytorch_bin,21730.99,22894.97,1264.96,0.0582,5,245.36,Tesla T4,t4,True
|
| 19 |
+
pythia-2.8b,EleutherAI/pythia-2.8b,pythia,small,2.8,5.294,1,2.03,checkpoint_format,pytorch_bin,23810.82,23824.29,11.02,0.0005,5,227.67,Tesla T4,t4,True
|
| 20 |
+
smollm3-3b,HuggingFaceTB/SmolLM3-3B,smollm3,small,3.08,5.728,2,1.997,checkpoint_format,pytorch_bin,25872.54,25885.11,17.18,0.0007,5,226.71,Tesla T4,t4,True
|
| 21 |
+
stablelm-2-1-6b,stabilityai/stablelm-2-1_6b-chat,stablelm,small,1.64,6.126,2,4.011,checkpoint_format,pytorch_bin,27647.71,27660.89,49.69,0.0018,5,226.89,Tesla T4,t4,True
|
| 22 |
+
bloomz-560m,bigscience/bloomz-560m,bloom,small,0.56,1.042,1,1.998,checkpoint_format,safetensors,695.19,918.93,105.77,0.1521,5,1534.85,Tesla T4,t4,True
|
| 23 |
+
qwen3-0.6b,Qwen/Qwen3-0.6B,qwen3,small,0.6,1.4,1,2.505,checkpoint_format,safetensors,949.5,1338.82,182.07,0.1918,5,1509.84,Tesla T4,t4,True
|
| 24 |
+
gpt2-medium,openai-community/gpt2-medium,gpt2,small,0.355,1.416,1,4.283,checkpoint_format,safetensors,890.56,1193.11,143.14,0.1607,5,1628.17,Tesla T4,t4,True
|
| 25 |
+
tinyllama-1.1b-v1.0,TinyLlama/TinyLlama-1.1B-Chat-v1.0,tinyllama,small,1.1,2.049,1,2.0,checkpoint_format,safetensors,1315.04,1755.38,214.39,0.163,5,1595.52,Tesla T4,t4,True
|
| 26 |
+
deepseek-coder-1.3b,deepseek-ai/deepseek-coder-1.3b-instruct,deepseek,small,1.35,2.508,1,1.995,checkpoint_format,safetensors,1715.57,5410.06,1865.45,1.0874,5,1496.99,Tesla T4,t4,False
|
| 27 |
+
pythia-1.4b,EleutherAI/pythia-1.4b,pythia,small,1.4,2.729,1,2.093,checkpoint_format,safetensors,1723.52,2185.96,226.97,0.1317,5,1621.39,Tesla T4,t4,True
|
| 28 |
+
qwen2.5-1.5b,Qwen/Qwen2.5-1.5B-Instruct,qwen2.5,small,1.54,2.875,1,2.005,checkpoint_format,safetensors,1807.28,2044.37,114.38,0.0633,5,1628.97,Tesla T4,t4,True
|
| 29 |
+
qwen2.5-coder-1.5b,Qwen/Qwen2.5-Coder-1.5B-Instruct,qwen2.5-coder,small,1.54,2.875,1,2.005,checkpoint_format,safetensors,2423.93,3570.22,725.26,0.2992,5,1214.56,Tesla T4,t4,True
|
| 30 |
+
gpt2-large,openai-community/gpt2-large,gpt2,small,0.774,3.024,1,4.195,checkpoint_format,safetensors,2128.77,4392.09,1099.29,0.5164,5,1454.63,Tesla T4,t4,False
|
| 31 |
+
falcon3-1b,tiiuae/Falcon3-1B-Instruct,falcon3,small,1.0,3.11,1,3.339,checkpoint_format,safetensors,2086.46,2794.37,366.85,0.1758,5,1526.34,Tesla T4,t4,True
|
| 32 |
+
smollm2-1.7b,HuggingFaceTB/SmolLM2-1.7B-Instruct,smollm2,small,1.71,3.188,1,2.002,checkpoint_format,safetensors,1980.48,2461.33,230.82,0.1165,5,1648.35,Tesla T4,t4,True
|
| 33 |
+
bloom-1b7,bigscience/bloom-1b7,bloom,small,1.7,3.208,1,2.026,checkpoint_format,safetensors,2017.37,7039.78,2380.52,1.18,5,1628.35,Tesla T4,t4,False
|
| 34 |
+
deepseek-r1-distill-qwen-1.5b,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,deepseek-r1,small,1.78,3.31,1,1.997,checkpoint_format,safetensors,2872.84,8002.05,2524.72,0.8788,5,1179.82,Tesla T4,t4,False
|
| 35 |
+
qwen3-1.7b,Qwen/Qwen3-1.7B,qwen3,small,1.7,3.784,2,2.39,checkpoint_format,safetensors,2394.41,3889.72,682.38,0.285,5,1618.28,Tesla T4,t4,True
|
| 36 |
+
granite-3.1-2b,ibm-granite/granite-3.1-2b-instruct,granite,small,2.53,4.719,2,2.003,checkpoint_format,safetensors,4069.39,6108.13,1083.3,0.2662,5,1187.46,Tesla T4,t4,True
|
| 37 |
+
phi-2,microsoft/phi-2,phi,small,2.78,5.178,2,2.0,checkpoint_format,safetensors,4854.01,16357.51,5425.72,1.1178,5,1092.35,Tesla T4,t4,False
|
| 38 |
+
qwen2.5-7b-awq,Qwen/Qwen2.5-7B-Instruct-AWQ,qwen2.5-awq,small,7.62,5.188,2,0.731,checkpoint_format,safetensors,4918.35,13317.7,3969.75,0.8071,5,1080.14,Tesla T4,t4,False
|
| 39 |
+
qwen2.5-7b-gptq-int4,Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4,qwen2.5-gptq,small,7.62,5.192,2,0.732,checkpoint_format,safetensors,5751.15,16538.7,4928.46,0.857,5,924.44,Tesla T4,t4,False
|
| 40 |
+
stablelm-zephyr-3b,stabilityai/stablelm-zephyr-3b,stablelm,small,2.8,5.207,1,1.997,checkpoint_format,safetensors,20583.48,21414.84,830.9,0.0404,5,259.04,Tesla T4,t4,True
|
| 41 |
+
pythia-2.8b,EleutherAI/pythia-2.8b,pythia,small,2.8,5.294,1,2.03,checkpoint_format,safetensors,6273.69,9939.74,2155.88,0.3436,5,864.09,Tesla T4,t4,False
|
| 42 |
+
smollm3-3b,HuggingFaceTB/SmolLM3-3B,smollm3,small,3.08,5.728,2,1.997,checkpoint_format,safetensors,7371.82,11363.7,2344.86,0.3181,5,795.66,Tesla T4,t4,False
|
| 43 |
+
stablelm-2-1-6b,stabilityai/stablelm-2-1_6b-chat,stablelm,small,1.64,6.126,2,4.011,checkpoint_format,safetensors,8701.54,14337.83,3662.89,0.4209,5,720.91,Tesla T4,t4,False
|
| 44 |
+
phi-3.5-mini,microsoft/Phi-3.5-mini-instruct,phi,small,3.82,7.117,2,2.0,checkpoint_format,safetensors,24946.31,27088.06,1088.52,0.0436,5,292.14,Tesla T4,t4,True
|
| 45 |
+
phi-3-mini-4k,microsoft/Phi-3-mini-4k-instruct,phi,small,3.82,7.117,2,2.0,checkpoint_format,safetensors,17158.02,27727.25,8742.37,0.5095,5,424.75,Tesla T4,t4,False
|
| 46 |
+
qwen3-4b,Qwen/Qwen3-4B,qwen3,small,4.0,7.492,3,2.011,checkpoint_format,safetensors,25288.99,25555.12,141.61,0.0056,5,303.37,Tesla T4,t4,True
|
| 47 |
+
bloomz-560m,bigscience/bloomz-560m,bloom,small,0.56,1.042,1,1.998,checkpoint_format,safetensors-nommap,685.26,689.58,8.12,0.0118,5,1557.08,Tesla T4,t4,True
|
| 48 |
+
qwen3-0.6b,Qwen/Qwen3-0.6B,qwen3,small,0.6,1.4,1,2.505,checkpoint_format,safetensors-nommap,925.59,1116.07,94.92,0.1025,5,1548.85,Tesla T4,t4,True
|
| 49 |
+
gpt2-medium,openai-community/gpt2-medium,gpt2,small,0.355,1.416,1,4.283,checkpoint_format,safetensors-nommap,912.33,1110.16,102.23,0.1121,5,1589.31,Tesla T4,t4,True
|
| 50 |
+
tinyllama-1.1b-v1.0,TinyLlama/TinyLlama-1.1B-Chat-v1.0,tinyllama,small,1.1,2.049,1,2.0,checkpoint_format,safetensors-nommap,1342.92,1737.58,186.82,0.1391,5,1562.39,Tesla T4,t4,True
|
| 51 |
+
deepseek-coder-1.3b,deepseek-ai/deepseek-coder-1.3b-instruct,deepseek,small,1.35,2.508,1,1.995,checkpoint_format,safetensors-nommap,1548.8,2045.2,226.28,0.1461,5,1658.18,Tesla T4,t4,True
|
| 52 |
+
pythia-1.4b,EleutherAI/pythia-1.4b,pythia,small,1.4,2.729,1,2.093,checkpoint_format,safetensors-nommap,1708.05,2306.55,299.02,0.1751,5,1636.08,Tesla T4,t4,True
|
| 53 |
+
qwen2.5-1.5b,Qwen/Qwen2.5-1.5B-Instruct,qwen2.5,small,1.54,2.875,1,2.005,checkpoint_format,safetensors-nommap,1794.08,2292.08,249.4,0.139,5,1640.96,Tesla T4,t4,True
|
| 54 |
+
qwen2.5-coder-1.5b,Qwen/Qwen2.5-Coder-1.5B-Instruct,qwen2.5-coder,small,1.54,2.875,1,2.005,checkpoint_format,safetensors-nommap,1814.78,2238.79,205.43,0.1132,5,1622.23,Tesla T4,t4,True
|
| 55 |
+
gpt2-large,openai-community/gpt2-large,gpt2,small,0.774,3.024,1,4.195,checkpoint_format,safetensors-nommap,1973.31,2334.45,171.86,0.0871,5,1569.23,Tesla T4,t4,True
|
| 56 |
+
falcon3-1b,tiiuae/Falcon3-1B-Instruct,falcon3,small,1.0,3.11,1,3.339,checkpoint_format,safetensors-nommap,1966.19,2539.21,282.26,0.1436,5,1619.7,Tesla T4,t4,True
|
| 57 |
+
smollm2-1.7b,HuggingFaceTB/SmolLM2-1.7B-Instruct,smollm2,small,1.71,3.188,1,2.002,checkpoint_format,safetensors-nommap,1991.74,2593.0,279.76,0.1405,5,1639.02,Tesla T4,t4,True
|
| 58 |
+
bloom-1b7,bigscience/bloom-1b7,bloom,small,1.7,3.208,1,2.026,checkpoint_format,safetensors-nommap,1951.27,2333.12,177.08,0.0907,5,1683.51,Tesla T4,t4,True
|
| 59 |
+
deepseek-r1-distill-qwen-1.5b,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,deepseek-r1,small,1.78,3.31,1,1.997,checkpoint_format,safetensors-nommap,2291.53,2885.34,269.63,0.1177,5,1479.12,Tesla T4,t4,True
|
| 60 |
+
qwen3-1.7b,Qwen/Qwen3-1.7B,qwen3,small,1.7,3.784,2,2.39,checkpoint_format,safetensors-nommap,2330.11,2908.15,290.51,0.1247,5,1662.93,Tesla T4,t4,True
|
| 61 |
+
granite-3.1-2b,ibm-granite/granite-3.1-2b-instruct,granite,small,2.53,4.719,2,2.003,checkpoint_format,safetensors-nommap,3981.05,4174.03,282.52,0.071,5,1213.82,Tesla T4,t4,True
|
| 62 |
+
phi-2,microsoft/phi-2,phi,small,2.78,5.178,2,2.0,checkpoint_format,safetensors-nommap,4535.22,5238.46,339.98,0.075,5,1169.13,Tesla T4,t4,True
|
| 63 |
+
qwen2.5-7b-awq,Qwen/Qwen2.5-7B-Instruct-AWQ,qwen2.5-awq,small,7.62,5.188,2,0.731,checkpoint_format,safetensors-nommap,5614.19,7816.06,1369.67,0.244,5,946.27,Tesla T4,t4,True
|
| 64 |
+
qwen2.5-7b-gptq-int4,Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4,qwen2.5-gptq,small,7.62,5.192,2,0.732,checkpoint_format,safetensors-nommap,6380.88,8036.84,1269.69,0.199,5,833.21,Tesla T4,t4,True
|
| 65 |
+
stablelm-zephyr-3b,stabilityai/stablelm-zephyr-3b,stablelm,small,2.8,5.207,1,1.997,checkpoint_format,safetensors-nommap,6279.41,13709.66,3620.08,0.5765,5,849.12,Tesla T4,t4,False
|
| 66 |
+
pythia-2.8b,EleutherAI/pythia-2.8b,pythia,small,2.8,5.294,1,2.03,checkpoint_format,safetensors-nommap,4392.32,5097.48,397.84,0.0906,5,1234.21,Tesla T4,t4,True
|
| 67 |
+
smollm3-3b,HuggingFaceTB/SmolLM3-3B,smollm3,small,3.08,5.728,2,1.997,checkpoint_format,safetensors-nommap,9030.49,11571.95,1420.31,0.1573,5,649.52,Tesla T4,t4,True
|
| 68 |
+
stablelm-2-1-6b,stabilityai/stablelm-2-1_6b-chat,stablelm,small,1.64,6.126,2,4.011,checkpoint_format,safetensors-nommap,7954.9,10001.57,1801.97,0.2265,5,788.57,Tesla T4,t4,True
|
| 69 |
+
phi-3.5-mini,microsoft/Phi-3.5-mini-instruct,phi,small,3.82,7.117,2,2.0,checkpoint_format,safetensors-nommap,11431.3,20865.54,5211.51,0.4559,5,637.53,Tesla T4,t4,False
|
| 70 |
+
phi-3-mini-4k,microsoft/Phi-3-mini-4k-instruct,phi,small,3.82,7.117,2,2.0,checkpoint_format,safetensors-nommap,8225.27,11290.91,2247.82,0.2733,5,886.03,Tesla T4,t4,True
|
| 71 |
+
qwen3-4b,Qwen/Qwen3-4B,qwen3,small,4.0,7.492,3,2.011,checkpoint_format,safetensors-nommap,23676.63,25215.73,4690.06,0.1981,5,324.02,Tesla T4,t4,True
|
| 72 |
+
bloomz-560m,bigscience/bloomz-560m,bloom,small,0.56,1.042,1,1.998,engine_init,transformers,2870.14,3227.81,241.06,0.084,3,371.76,Tesla T4,t4,True
|
| 73 |
+
qwen3-0.6b,Qwen/Qwen3-0.6B,qwen3,small,0.6,1.4,1,2.505,engine_init,transformers,3109.51,7340.83,2553.73,0.8213,3,461.04,Tesla T4,t4,False
|
| 74 |
+
gpt2-medium,openai-community/gpt2-medium,gpt2,small,0.355,1.416,1,4.283,engine_init,transformers,1285.02,5730.36,2405.16,1.8717,3,1128.37,Tesla T4,t4,False
|
| 75 |
+
tinyllama-1.1b-v1.0,TinyLlama/TinyLlama-1.1B-Chat-v1.0,tinyllama,small,1.1,2.049,1,2.0,engine_init,transformers,3825.68,6858.9,2452.1,0.641,3,548.45,Tesla T4,t4,False
|
| 76 |
+
deepseek-coder-1.3b,deepseek-ai/deepseek-coder-1.3b-instruct,deepseek,small,1.35,2.508,1,1.995,engine_init,transformers,1947.16,7655.35,3137.44,1.6113,3,1318.95,Tesla T4,t4,False
|
| 77 |
+
pythia-1.4b,EleutherAI/pythia-1.4b,pythia,small,1.4,2.729,1,2.093,engine_init,transformers,1334.53,3878.42,1347.07,1.0094,3,2094.0,Tesla T4,t4,False
|
| 78 |
+
qwen2.5-1.5b,Qwen/Qwen2.5-1.5B-Instruct,qwen2.5,small,1.54,2.875,1,2.005,engine_init,transformers,2262.23,5195.45,1542.23,0.6817,3,1301.37,Tesla T4,t4,False
|
| 79 |
+
qwen2.5-coder-1.5b,Qwen/Qwen2.5-Coder-1.5B-Instruct,qwen2.5-coder,small,1.54,2.875,1,2.005,engine_init,transformers,4861.99,10770.98,3882.37,0.7985,3,605.51,Tesla T4,t4,False
|
| 80 |
+
gpt2-large,openai-community/gpt2-large,gpt2,small,0.774,3.024,1,4.195,engine_init,transformers,4213.05,10072.76,3759.56,0.8924,3,735.0,Tesla T4,t4,False
|
| 81 |
+
falcon3-1b,tiiuae/Falcon3-1B-Instruct,falcon3,small,1.0,3.11,1,3.339,engine_init,transformers,3882.89,12556.54,4954.31,1.2759,3,820.17,Tesla T4,t4,False
|
| 82 |
+
smollm2-1.7b,HuggingFaceTB/SmolLM2-1.7B-Instruct,smollm2,small,1.71,3.188,1,2.002,engine_init,transformers,1519.97,7757.82,3274.13,2.1541,3,2147.74,Tesla T4,t4,False
|
| 83 |
+
bloom-1b7,bigscience/bloom-1b7,bloom,small,1.7,3.208,1,2.026,engine_init,transformers,10647.37,10798.58,2960.13,0.278,3,308.53,Tesla T4,t4,True
|
| 84 |
+
deepseek-r1-distill-qwen-1.5b,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,deepseek-r1,small,1.78,3.31,1,1.997,engine_init,transformers,3525.12,13968.93,5653.96,1.6039,3,961.51,Tesla T4,t4,False
|
| 85 |
+
qwen3-1.7b,Qwen/Qwen3-1.7B,qwen3,small,1.7,3.784,2,2.39,engine_init,transformers,8692.28,18968.11,7204.68,0.8289,3,445.78,Tesla T4,t4,False
|
| 86 |
+
granite-3.1-2b,ibm-granite/granite-3.1-2b-instruct,granite,small,2.53,4.719,2,2.003,engine_init,transformers,12946.29,19295.65,6893.92,0.5325,3,373.25,Tesla T4,t4,False
|
| 87 |
+
phi-2,microsoft/phi-2,phi,small,2.78,5.178,2,2.0,engine_init,transformers,17282.65,19939.73,6935.56,0.4013,3,306.8,Tesla T4,t4,False
|
| 88 |
+
stablelm-zephyr-3b,stabilityai/stablelm-zephyr-3b,stablelm,small,2.8,5.207,1,1.997,engine_init,transformers,19660.7,20809.66,3447.14,0.1753,3,271.2,Tesla T4,t4,True
|
| 89 |
+
pythia-2.8b,EleutherAI/pythia-2.8b,pythia,small,2.8,5.294,1,2.03,engine_init,transformers,19414.53,21810.34,3112.51,0.1603,3,279.23,Tesla T4,t4,True
|
| 90 |
+
smollm3-3b,HuggingFaceTB/SmolLM3-3B,smollm3,small,3.08,5.728,2,1.997,engine_init,transformers,22397.24,23205.72,5126.65,0.2289,3,261.88,Tesla T4,t4,True
|
| 91 |
+
stablelm-2-1-6b,stabilityai/stablelm-2-1_6b-chat,stablelm,small,1.64,6.126,2,4.011,engine_init,transformers,19685.11,21764.45,1300.91,0.0661,3,318.67,Tesla T4,t4,True
|
| 92 |
+
phi-3.5-mini,microsoft/Phi-3.5-mini-instruct,phi,small,3.82,7.117,2,2.0,engine_init,transformers,25219.89,30470.21,3045.41,0.1208,3,288.97,Tesla T4,t4,True
|
| 93 |
+
phi-3-mini-4k,microsoft/Phi-3-mini-4k-instruct,phi,small,3.82,7.117,2,2.0,engine_init,transformers,15268.86,22811.14,6656.83,0.436,3,477.3,Tesla T4,t4,False
|
| 94 |
+
qwen3-4b,Qwen/Qwen3-4B,qwen3,small,4.0,7.492,3,2.011,engine_init,transformers,32412.47,32641.33,250.79,0.0077,3,236.69,Tesla T4,t4,True
|
| 95 |
+
bloomz-560m,bigscience/bloomz-560m,bloom,small,0.56,1.042,1,1.998,storage_tier,local-nvme,678.07,776.71,53.97,0.0796,5,1573.59,Tesla T4,t4,True
|
| 96 |
+
qwen3-0.6b,Qwen/Qwen3-0.6B,qwen3,small,0.6,1.4,1,2.505,storage_tier,local-nvme,1054.55,1278.35,146.51,0.1389,5,1359.44,Tesla T4,t4,True
|
| 97 |
+
gpt2-medium,openai-community/gpt2-medium,gpt2,small,0.355,1.416,1,4.283,storage_tier,local-nvme,922.55,928.6,7.5,0.0081,5,1571.72,Tesla T4,t4,True
|
| 98 |
+
tinyllama-1.1b-v1.0,TinyLlama/TinyLlama-1.1B-Chat-v1.0,tinyllama,small,1.1,2.049,1,2.0,storage_tier,local-nvme,1367.22,1845.32,241.16,0.1764,5,1534.62,Tesla T4,t4,True
|
| 99 |
+
deepseek-coder-1.3b,deepseek-ai/deepseek-coder-1.3b-instruct,deepseek,small,1.35,2.508,1,1.995,storage_tier,local-nvme,1567.02,3173.98,800.53,0.5109,5,1638.91,Tesla T4,t4,False
|
| 100 |
+
pythia-1.4b,EleutherAI/pythia-1.4b,pythia,small,1.4,2.729,1,2.093,storage_tier,local-nvme,1744.79,3175.54,711.31,0.4077,5,1601.62,Tesla T4,t4,False
|
| 101 |
+
qwen2.5-1.5b,Qwen/Qwen2.5-1.5B-Instruct,qwen2.5,small,1.54,2.875,1,2.005,storage_tier,local-nvme,2113.14,2746.62,387.79,0.1835,5,1393.19,Tesla T4,t4,True
|
| 102 |
+
qwen2.5-coder-1.5b,Qwen/Qwen2.5-Coder-1.5B-Instruct,qwen2.5-coder,small,1.54,2.875,1,2.005,storage_tier,local-nvme,2338.3,4674.72,1267.91,0.5422,5,1259.04,Tesla T4,t4,False
|
| 103 |
+
gpt2-large,openai-community/gpt2-large,gpt2,small,0.774,3.024,1,4.195,storage_tier,local-nvme,1908.99,2512.38,280.1,0.1467,5,1622.1,Tesla T4,t4,True
|
| 104 |
+
falcon3-1b,tiiuae/Falcon3-1B-Instruct,falcon3,small,1.0,3.11,1,3.339,storage_tier,local-nvme,2097.43,2512.33,255.75,0.1219,5,1518.35,Tesla T4,t4,True
|
| 105 |
+
smollm2-1.7b,HuggingFaceTB/SmolLM2-1.7B-Instruct,smollm2,small,1.71,3.188,1,2.002,storage_tier,local-nvme,2180.12,2618.96,279.76,0.1283,5,1497.4,Tesla T4,t4,True
|
| 106 |
+
bloom-1b7,bigscience/bloom-1b7,bloom,small,1.7,3.208,1,2.026,storage_tier,local-nvme,2055.54,2637.45,296.26,0.1441,5,1598.12,Tesla T4,t4,True
|
| 107 |
+
deepseek-r1-distill-qwen-1.5b,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,deepseek-r1,small,1.78,3.31,1,1.997,storage_tier,local-nvme,2132.53,2457.43,162.12,0.076,5,1589.39,Tesla T4,t4,True
|
| 108 |
+
qwen3-1.7b,Qwen/Qwen3-1.7B,qwen3,small,1.7,3.784,2,2.39,storage_tier,local-nvme,2967.42,8231.97,2597.01,0.8752,5,1305.79,Tesla T4,t4,False
|
| 109 |
+
phi-2,microsoft/phi-2,phi,small,2.78,5.178,2,2.0,storage_tier,local-nvme,9390.04,21252.69,6516.98,0.694,5,564.67,Tesla T4,t4,False
|
| 110 |
+
qwen2.5-7b-awq,Qwen/Qwen2.5-7B-Instruct-AWQ,qwen2.5-awq,small,7.62,5.188,2,0.731,storage_tier,local-nvme,5635.85,15885.83,5272.93,0.9356,5,942.63,Tesla T4,t4,False
|
| 111 |
+
qwen2.5-7b-gptq-int4,Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4,qwen2.5-gptq,small,7.62,5.192,2,0.732,storage_tier,local-nvme,5151.85,13855.94,4192.54,0.8138,5,1031.98,Tesla T4,t4,False
|
| 112 |
+
stablelm-zephyr-3b,stabilityai/stablelm-zephyr-3b,stablelm,small,2.8,5.207,1,1.997,storage_tier,local-nvme,21967.62,21987.86,27.13,0.0012,5,242.72,Tesla T4,t4,True
|
| 113 |
+
pythia-2.8b,EleutherAI/pythia-2.8b,pythia,small,2.8,5.294,1,2.03,storage_tier,local-nvme,22252.33,22310.64,34.53,0.0016,5,243.62,Tesla T4,t4,True
|
| 114 |
+
smollm3-3b,HuggingFaceTB/SmolLM3-3B,smollm3,small,3.08,5.728,2,1.997,storage_tier,local-nvme,8822.38,17518.78,5179.79,0.5871,5,664.84,Tesla T4,t4,False
|
| 115 |
+
stablelm-2-1-6b,stabilityai/stablelm-2-1_6b-chat,stablelm,small,1.64,6.126,2,4.011,storage_tier,local-nvme,25999.38,26030.78,61.27,0.0024,5,241.28,Tesla T4,t4,True
|
| 116 |
+
phi-3.5-mini,microsoft/Phi-3.5-mini-instruct,phi,small,3.82,7.117,2,2.0,storage_tier,local-nvme,30168.3,30227.47,32.57,0.0011,5,241.57,Tesla T4,t4,True
|
| 117 |
+
phi-3-mini-4k,microsoft/Phi-3-mini-4k-instruct,phi,small,3.82,7.117,2,2.0,storage_tier,local-nvme,30136.45,30206.59,37.92,0.0013,5,241.83,Tesla T4,t4,True
|
| 118 |
+
qwen3-4b,Qwen/Qwen3-4B,qwen3,small,4.0,7.492,3,2.011,storage_tier,local-nvme,27124.15,27317.04,111.12,0.0041,5,282.84,Tesla T4,t4,True
|
| 119 |
+
bloomz-560m,bigscience/bloomz-560m,bloom,small,0.56,1.042,1,1.998,storage_tier,remote-emulated-200MiBs,6011.45,8490.07,1227.43,0.2042,5,177.5,Tesla T4,t4,True
|
| 120 |
+
qwen3-0.6b,Qwen/Qwen3-0.6B,qwen3,small,0.6,1.4,1,2.505,storage_tier,remote-emulated-200MiBs,8105.65,8964.32,429.4,0.053,5,176.86,Tesla T4,t4,True
|
| 121 |
+
gpt2-medium,openai-community/gpt2-medium,gpt2,small,0.355,1.416,1,4.283,storage_tier,remote-emulated-200MiBs,8176.08,8447.3,138.5,0.0169,5,177.34,Tesla T4,t4,True
|
| 122 |
+
tinyllama-1.1b-v1.0,TinyLlama/TinyLlama-1.1B-Chat-v1.0,tinyllama,small,1.1,2.049,1,2.0,storage_tier,remote-emulated-200MiBs,11788.49,12686.29,407.88,0.0346,5,177.99,Tesla T4,t4,True
|
| 123 |
+
deepseek-coder-1.3b,deepseek-ai/deepseek-coder-1.3b-instruct,deepseek,small,1.35,2.508,1,1.995,storage_tier,remote-emulated-200MiBs,14795.63,14853.32,202.8,0.0137,5,173.58,Tesla T4,t4,True
|
| 124 |
+
pythia-1.4b,EleutherAI/pythia-1.4b,pythia,small,1.4,2.729,1,2.093,storage_tier,remote-emulated-200MiBs,15883.07,16364.86,288.0,0.0181,5,175.94,Tesla T4,t4,True
|
| 125 |
+
qwen2.5-1.5b,Qwen/Qwen2.5-1.5B-Instruct,qwen2.5,small,1.54,2.875,1,2.005,storage_tier,remote-emulated-200MiBs,17075.09,20699.8,1762.22,0.1032,5,172.41,Tesla T4,t4,True
|
| 126 |
+
qwen2.5-coder-1.5b,Qwen/Qwen2.5-Coder-1.5B-Instruct,qwen2.5-coder,small,1.54,2.875,1,2.005,storage_tier,remote-emulated-200MiBs,16745.75,17474.0,441.23,0.0263,5,175.81,Tesla T4,t4,True
|
| 127 |
+
gpt2-large,openai-community/gpt2-large,gpt2,small,0.774,3.024,1,4.195,storage_tier,remote-emulated-200MiBs,17375.51,18597.08,549.0,0.0316,5,178.22,Tesla T4,t4,True
|
| 128 |
+
falcon3-1b,tiiuae/Falcon3-1B-Instruct,falcon3,small,1.0,3.11,1,3.339,storage_tier,remote-emulated-200MiBs,17963.8,19028.87,556.02,0.031,5,177.28,Tesla T4,t4,True
|
| 129 |
+
smollm2-1.7b,HuggingFaceTB/SmolLM2-1.7B-Instruct,smollm2,small,1.71,3.188,1,2.002,storage_tier,remote-emulated-200MiBs,19126.9,19606.29,312.0,0.0163,5,170.68,Tesla T4,t4,True
|
| 130 |
+
bloom-1b7,bigscience/bloom-1b7,bloom,small,1.7,3.208,1,2.026,storage_tier,remote-emulated-200MiBs,18557.8,19043.75,251.84,0.0136,5,177.01,Tesla T4,t4,True
|
| 131 |
+
deepseek-r1-distill-qwen-1.5b,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,deepseek-r1,small,1.78,3.31,1,1.997,storage_tier,remote-emulated-200MiBs,19013.22,20817.13,873.88,0.046,5,178.27,Tesla T4,t4,True
|
| 132 |
+
qwen3-1.7b,Qwen/Qwen3-1.7B,qwen3,small,1.7,3.784,2,2.39,storage_tier,remote-emulated-200MiBs,22459.24,26589.9,2207.02,0.0983,5,172.53,Tesla T4,t4,True
|
| 133 |
+
phi-2,microsoft/phi-2,phi,small,2.78,5.178,2,2.0,storage_tier,remote-emulated-200MiBs,32540.47,45733.33,6586.88,0.2024,5,162.94,Tesla T4,t4,True
|
| 134 |
+
qwen2.5-7b-awq,Qwen/Qwen2.5-7B-Instruct-AWQ,qwen2.5-awq,small,7.62,5.188,2,0.731,storage_tier,remote-emulated-200MiBs,31946.74,41997.27,4964.89,0.1554,5,166.29,Tesla T4,t4,True
|
| 135 |
+
qwen2.5-7b-gptq-int4,Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4,qwen2.5-gptq,small,7.62,5.192,2,0.732,storage_tier,remote-emulated-200MiBs,31566.65,41264.7,4875.4,0.1544,5,168.42,Tesla T4,t4,True
|
| 136 |
+
stablelm-zephyr-3b,stabilityai/stablelm-zephyr-3b,stablelm,small,2.8,5.207,1,1.997,storage_tier,remote-emulated-200MiBs,48714.58,48766.47,35.75,0.0007,5,109.45,Tesla T4,t4,True
|
| 137 |
+
pythia-2.8b,EleutherAI/pythia-2.8b,pythia,small,2.8,5.294,1,2.03,storage_tier,remote-emulated-200MiBs,49399.74,49462.14,107.2,0.0022,5,109.74,Tesla T4,t4,True
|
| 138 |
+
smollm3-3b,HuggingFaceTB/SmolLM3-3B,smollm3,small,3.08,5.728,2,1.997,storage_tier,remote-emulated-200MiBs,40280.15,49405.74,5869.1,0.1457,5,145.62,Tesla T4,t4,True
|
| 139 |
+
stablelm-2-1-6b,stabilityai/stablelm-2-1_6b-chat,stablelm,small,1.64,6.126,2,4.011,storage_tier,remote-emulated-200MiBs,57305.97,57339.69,16.81,0.0003,5,109.47,Tesla T4,t4,True
|
| 140 |
+
phi-3.5-mini,microsoft/Phi-3.5-mini-instruct,phi,small,3.82,7.117,2,2.0,storage_tier,remote-emulated-200MiBs,63060.84,66701.68,4687.84,0.0743,5,115.57,Tesla T4,t4,True
|
| 141 |
+
phi-3-mini-4k,microsoft/Phi-3-mini-4k-instruct,phi,small,3.82,7.117,2,2.0,storage_tier,remote-emulated-200MiBs,66585.73,66600.89,15.78,0.0002,5,109.45,Tesla T4,t4,True
|
| 142 |
+
qwen3-4b,Qwen/Qwen3-4B,qwen3,small,4.0,7.492,3,2.011,storage_tier,remote-emulated-200MiBs,66268.19,66551.85,150.49,0.0023,5,115.77,Tesla T4,t4,True
|
runs.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|