Ben Graville commited on
README: ctx8192 bundles
Browse files
README.md
CHANGED
|
@@ -27,3 +27,14 @@ below; all original model licences apply and travel with these weights.
|
|
| 27 |
- Bundles carry `runtime_env COREAI_CHUNK_THRESHOLD=1`.
|
| 28 |
|
| 29 |
Upstream sources: `google/gemma-4-E4B-it-qat-q4_0-unquantized`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
- Bundles carry `runtime_env COREAI_CHUNK_THRESHOLD=1`.
|
| 28 |
|
| 29 |
Upstream sources: `google/gemma-4-E4B-it-qat-q4_0-unquantized`
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## High-context bundles — ctx 8192
|
| 33 |
+
|
| 34 |
+
Same weights, same graph, same recipe as the 4096 bundles above — the exporter's
|
| 35 |
+
`--max-ctx` only widens the dynamic sequence bound and the manifest's
|
| 36 |
+
`language.max_context_length`. KV cache stays fp16.
|
| 37 |
+
|
| 38 |
+
| Bundle | Source checkpoint | Context | Functions | KV dtype | Producer | Created |
|
| 39 |
+
|---|---|---|---|---|---|---|
|
| 40 |
+
| `ctx8192/gpu-pipelined/gemma4_e4b_qat_decode_int4lin_tbl_pf64` | google/gemma-4-E4B-it-qat-q4_0-unquantized | 8192 | main+prefill | fp16 | coreai-core 1.0.0b2 | 20260817T204459Z |
|