Document qualified RTX PRO 6000 adaptive-MTP source and benchmark results
Browse files
README.md
CHANGED
|
@@ -54,6 +54,46 @@ Sidecar SHA-256:
|
|
| 54 |
|
| 55 |
## Performance
|
| 56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
### RTX PRO 6000 Blackwell
|
| 58 |
|
| 59 |
Matched tests on one RTX PRO 6000 Blackwell used native MTP, CUDA graphs, and a
|
|
|
|
| 54 |
|
| 55 |
## Performance
|
| 56 |
|
| 57 |
+
### September 2026: qualified RTX PRO adaptive-MTP results
|
| 58 |
+
|
| 59 |
+
A separate, versioned [runtime source bundle](https://github.com/garnermccloud/sglang-ssd-stream/tree/rtx-pro-6000-20260909/optimizations/rtx-pro-6000-20260909)
|
| 60 |
+
contains the exact optimization overlay deployed on September 9, including
|
| 61 |
+
corrected sparse-index packing, adaptive four/eight-token MTP windows, corrected
|
| 62 |
+
ReplaySSM, shared scratch/capture resources, and early draft-head initialization.
|
| 63 |
+
**The model weights did not change.**
|
| 64 |
+
|
| 65 |
+
Matched measurements on one RTX PRO 6000 Blackwell Workstation Edition (96 GB):
|
| 66 |
+
|
| 67 |
+
| Workload | Fixed-four tok/s | Adaptive bundle tok/s | Median change |
|
| 68 |
+
| --- | ---: | ---: | ---: |
|
| 69 |
+
| List | 175.57 | 186.13 | +6.01% |
|
| 70 |
+
| Prose | 187.64 | 194.61 | +3.71% |
|
| 71 |
+
| Code | 238.76 | 282.12 | +18.16% |
|
| 72 |
+
| Reasoning | 163.35 | 163.97 | +0.38% |
|
| 73 |
+
|
| 74 |
+
Both arms used corrected packing, original weights, strict acceptance thresholds
|
| 75 |
+
of 1, FP32 SSM, FP8 KV, CUDA graphs, and a 262,144-token context/pool. Each
|
| 76 |
+
arm/workload has six measured 1,024-token completions: 48 total measurements,
|
| 77 |
+
excluding 16 warmups. The source bundle includes file hashes, a readable patch,
|
| 78 |
+
all timing records, input token IDs, synthetic prompts, and an offline verifier.
|
| 79 |
+
The historical qualification passed 12 functional checks and five GPU tests.
|
| 80 |
+
|
| 81 |
+
This is a workload-specific code-generation gain, **not a universal 18% speedup
|
| 82 |
+
or a coding-quality claim**. Outputs varied, and the small non-code differences
|
| 83 |
+
remain sensitive to output and run order. One candidate code request measured
|
| 84 |
+
241.15 tok/s. The historical runtime used the original RadixArk checkpoint and
|
| 85 |
+
legacy SSD integration, not a fresh installation of this prepared model page.
|
| 86 |
+
|
| 87 |
+
**The normal installer below does not enable this bundle.** Its pinned SGLang
|
| 88 |
+
revision and plugin interface differ from the qualified historical runtime.
|
| 89 |
+
Porting and clean-install acceptance are still pending; the published overlay
|
| 90 |
+
must not be copied into arbitrary SGLang installations. No complete legacy base
|
| 91 |
+
image is included. See the [versioned publication](https://github.com/garnermccloud/sglang-ssd-stream/releases/tag/rtx-pro-6000-20260909)
|
| 92 |
+
for applicability, verification commands, and test limitations.
|
| 93 |
+
|
| 94 |
+
These results are a separate experiment and should not be compared directly
|
| 95 |
+
with the earlier SSD-versus-RAM figures below.
|
| 96 |
+
|
| 97 |
### RTX PRO 6000 Blackwell
|
| 98 |
|
| 99 |
Matched tests on one RTX PRO 6000 Blackwell used native MTP, CUDA graphs, and a
|