8b cap bundle: README with provenance, KV arithmetic, digests

aff4-c128000/README.md

@@ -0,0 +1,67 @@
+# LFM2.5 8B-A1B — aff4 D2 at the checkpoint's maximum context (128,000)
+
+Core AI bundle for Apple Silicon Macs. **Byte-for-byte the `aff4/` D2 recipe** — affine-int4
+experts (K-block 64, gather_qmm kernel), lm_head untied and quantized int8 **symmetric**
+per-block-32, attention matmul fp16 — with exactly one argument changed:
+`--max-ctx 128000` in place of `16384`.
+
+## Why this is one bundle and not a tier ladder
+
+The LFM2.5 export declares its KV sequence dimension as a **dynamic** `torch.export.Dim`, not a
+pinned extent. The saved asset says so:
+
+    main  ST  keyCache   Float16  6 x 1 x 8 x ? x 64
+    main  ST  valueCache Float16  6 x 1 x 8 x ? x 64
+    main  ST  convState  Float16  18 x 1 x 2048 x 2
+
+Those descriptors are **character-identical** to the cap-16,384 build. At runtime
+`KVCacheFactory` resolves `.auto` to `GrowingKVCache`, whose `defaultSize` is the constant
+**256 tokens** regardless of the declared cap; it doubles on demand. So raising the cap
+allocates nothing — memory scales with use, not with declaration.
+
+Measured cost of the change: `main.mlirb` moved **+22 bytes** (5,707,008,966 -> 5,707,008,988),
+and the first differing byte is at offset 5,706,962,354 — **99.99918% of the file is byte
+identical**, every constant included. For reference, running the identical export command twice
+differs by ~22,000 bytes. The weights are the D2 weights; the fp32-oracle rows cannot move.
+
+## 128,000, not 131,072
+
+`config.json` gives `max_position_embeddings = 128000` with `rope_type "default"`,
+`rope_theta 5e6` and **no scaling factor**. 131,072 would be 3,072 positions past everything the
+checkpoint declares. This bundle stops where the checkpoint does.
+
+## What a context window costs
+
+18 of 24 layers are short-convolution mixers whose entire state is `convState` — **147,456
+bytes, fixed, capacity-invariant**. All long-range state rides 6 GQA layers at
+`6 x 8 kv-heads x 64 head_dim x 2 B x 2` = **12,288 B/token**.
+
+| window | KV allocated (`GrowingKVCache` rounds to a power of two) |
+|---:|---:|
+| 4,096 | 0.047 GiB |
+| 16,384 | 0.188 GiB |
+| 65,536 | 0.750 GiB |
+| 128,000 | 1.500 GiB (131,072 slots) |
+
+The whole ladder from 4k to the checkpoint maximum spans 1.45 GiB.
+
+## Files and sha256
+
+    metadata.json                       647 B  e00a70e5820db09928bc338efc01e6abd6f2d7ca19b6f2ff062c202ceb2a7d39
+    *.aimodel/main.hash                  32 B  7b999d0d329b497afab466891180638116181971c6fe19c956c116cf65b0db1c
+    *.aimodel/main.mlirb       5707008988 B    446df156d77cc1b5d2a54c49bb573930cbe22d0281bd7e16f86f24e5da55a663
+    *.aimodel/metadata.json             105 B  bea3d473ac849d482e4ef338dcdafc1cc8b3db833b07494f903a339a4fe5bd23
+    tokenizer/chat_template.jinja      4779 B  b2881deb506bdff56bf6ac46aaa33292c119953f82fedaf276634fcc4282a559
+    tokenizer/tokenizer.json       17905598 B  695be7802a0e4b8a81048f0ff5ebb7fc811a0ba5a6be63dbb24deb5a81096f41
+    tokenizer/tokenizer_config.json     315 B  b53c3caceb7fa0de424e1aa67b3cc5008cd2f8b8d2426a8fa221ef3cf1a949d5
+    Total: 5,724,920,464 bytes
+
+The three tokenizer files are byte-identical to the pinned `aff4/` D2 bundle.
+
+Built 2026-08-27, `coreai-torch 0.4.1` / `coreai-core 1.0.0b2`, producer
+`coreai-core 1.0.0b2`, rc=0, 294 s, peak export RSS 46.74 GB.
+
+**Not yet compiled or run on a Mac.** No token has been generated from this bundle at any
+context.
+
+License: LFM Open License v1.0, inherited from LiquidAI/LFM2.5-8B-A1B (see LICENSE at repo root).