# GLM-5.2 EXL3 release validation

These artifacts were collected on four RTX PRO 6000 Blackwell 96 GB GPUs with
the final release image, TP4 + DCP4 A2A, MTP1 greedy, asynchronous scheduling
disabled, and NVFP4 DeepSeek-MLA KV cache unless a row says otherwise.

Tested local image ID:

```text
sha256:bfd6d6670db37b04e9cbef7375722e3f71d66745abf1714c05cc5b71fd126715
```

Published immutable image:

```text
verdictai/glm52-exl3-sparkinfer:v2-gg-1043999ab-spi879ca0ad-cu132-sm120a@sha256:bfd6d6670db37b04e9cbef7375722e3f71d66745abf1714c05cc5b71fd126715
```

Source revisions:

- Gilded Gnosis vLLM: `1043999ab5f13350aaacc188e713d6c0928387e9`
- Sparkinfer: `879ca0ad878958a0fbb1d6a1393c1bb40ec54790`

The vLLM banner in the logs retains the base package's `g60c82d972` version
string. The `ai.verdict.vllm.revision` and `ai.verdict.sparkinfer.revision` OCI
labels in `image-inspect.json` identify the source actually installed in the
tested image.

## Correctness

| Evaluation | Result | Concurrency | Max output | Cap hits |
| --- | ---: | ---: | ---: | ---: |
| LAVD pass 1 | 10/10 (5 exact, 5 near) | 5 | 20,000 | 0 |
| LAVD pass 2 | 10/10 (3 exact, 7 near) | 5 | 20,000 | 0 |
| Estonia | 5/5 pass | 5 | 5,000 | 0 |

The two LAVD passes are independent full runs, for a combined 20/20 scorer
pass. `near` is a passing LAVD score, not a failure.
LAVD used temperature 0 and repetition penalty 1.15; Estonia used temperature
0 and repetition penalty 1.25.

## KV-cache KLD

Each KLD figure is the mean of five independent runs against the same verified
BF16 reference-logit artifact. Every run evaluated all 2,047 positions from a
2,048-token sample with TP4 + DCP4 A2A, the 78-layer sparse-index pattern,
asynchronous scheduling disabled, and eager execution. KLD isolates model and
KV-cache drift; the serving evaluations above validate MTP1 and CUDA graphs.
The reference artifact SHA256 is
`87f992a689c054a0548a4b3863da6c809f9239beacd5786d0401e45904fec063`.

| KV cache | Mean KLD | Sample SD | Min | Max | Runs |
| --- | ---: | ---: | ---: | ---: | ---: |
| NVFP4 DeepSeek MLA | 0.1124021 | 0.0025948 | 0.1086084 | 0.1156108 | 5 |
| FP8 | 0.1036666 | 0.0018374 | 0.1016535 | 0.1066077 | 5 |

## Prefill

Cold standalone prefill uses `prompt_tokens / TTFT`; the client figure is the
headline and the uncontaminated vLLM Prometheus counter is shown as validation.

| Requested context | Prompt tokens | TTFT | Client tok/s | Server tok/s |
| ---: | ---: | ---: | ---: | ---: |
| 8K | 8,201 | 5.46 s | 1,502 | 1,507 |
| 64K | 64,512 | 51.64 s | 1,249 | 1,252 |
| 128K | 128,881 | 109.00 s | 1,182 | 1,185 |

## Sustained decode

These are 20-second steady-state cells after warmup, at zero input context and
with EOS ignored. Aggregate throughput comes from continuous OpenAI stream
usage. There were no request errors, underfilled cells, or capacity-limited
cells. A separate 30-second C1 run measured 48.5 tok/s.

| Concurrency | Aggregate tok/s | Per-request tok/s |
| ---: | ---: | ---: |
| 1 | 48.9 | 48.9 |
| 2 | 112.9 | 56.4 |
| 3 | 154.0 | 51.3 |
| 4 | 188.2 | 47.0 |
| 5 | 218.2 | 43.6 |
| 6 | 239.4 | 39.9 |
| 7 | 253.9 | 36.3 |
| 8 | 266.8 | 33.3 |

The release preset allocates 524,288 logical KV-cache tokens: 2,048 blocks x
64 tokens x DCP4, with 131,072 tokens local to each DCP rank. This is the
validated allocation and configured per-request context cap.

## Raw artifacts

- [C1-C8 decode JSON](decode-c1-c8.json) and [Rich TUI log](decode-c1-c8.log)
- [Dedicated C1 JSON](decode-c1-dedicated.json) and [Rich TUI log](decode-c1-dedicated.log)
- [Prefill JSON](prefill-8k-64k-128k.json) and [Rich TUI log](prefill-8k-64k-128k.log)
- [LAVD pass 1 JSON](lavd-c5-r10.json) and [Rich TUI log](lavd-c5-r10.log)
- [LAVD pass 2 JSON](lavd-c5-r10-pass2.json) and [Rich TUI log](lavd-c5-r10-pass2.log)
- [Estonia JSON](estonia-c5-r5.json) and [Rich TUI log](estonia-c5-r5.log)
- [NVFP4 KLD summary](kld-nvfp4-dcp4-summary.json)
- [FP8 KLD summary](kld-fp8-dcp4-summary.json)
- [Exact tested-image inspection](image-inspect.json)

At C8, the hottest observed GPU reached 90 C. NVIDIA hardware and software
thermal-slowdown flags were inactive during the run.