Document split body.vmfb + lm_head.vmfb alongside monolithic model.vmfb

README.md

@@ -16,27 +16,40 @@ pipeline_tag: text-generation
 # LFM2.5-230M — Torq build (Synaptics SL2619 NPU)
 
 Pre-compiled **Torq VMFB** build of LiquidAI's LFM2.5 230M text language model, ready
-to run on the Synaptics **SL2619** edge NPU. The model runs on the
+to run on the Synaptics **SL2619** edge NPU. LFM2 is a hybrid architecture that
+combines short convolutions with grouped-query attention. The transformer runs on the
 NPU in bf16; the token embeddings run on the host CPU.
 
 ## Contents
 
 | File | Size | Role |
 |---|--:|---|
-| `model.vmfb` | 461 MB | LFM2.5-230M decoder + LM head, compiled for the Torq NPU |
+| `model.vmfb` | 461 MB | **monolithic** build — decoder + LM head in one graph (logits output) |
+| `body.vmfb` | 327 MB | **split** build — decoder body only, emits hidden states (pairs with `lm_head.vmfb`) |
+| `lm_head.vmfb` | 134 MB | **split** build — standalone LM head (hidden → 65 536 logits) |
 | `token_embeddings.npy` | 134 MB | CPU embedding lookup table (bf16) |
 | `config.json` | — | model configuration |
 | `tokenizer.json`, `tokenizer_config.json` | — | tokenizer + tokenizer config |
 | `onnx/model.onnx` (+ `model.onnx_data`) | ~952 MB | reference ONNX export for non-Torq runtimes (e.g. onnxruntime) |
 
-The `.vmfb` is the artifact to run on the Torq NPU; the `onnx/` export is provided for
-reference / portability to other runtimes.
+### Monolithic vs. split
+
+Two equivalent ways to run the model (same weights — `body` 327 MB + `lm_head` 134 MB ≈
+the 461 MB monolithic build):
+
+- **`model.vmfb` (monolithic):** one graph that outputs logits directly. Simplest to run.
+- **`body.vmfb` + `lm_head.vmfb` (split):** the decoder body emits hidden states and the
+  LM head is applied only when sampling. Prefill tokens then skip the large
+  `[1024 → 65 536]` LM-head projection, which **lowers time-to-first-token** — pick this
+  when TTFT matters.
+
+The `onnx/` export is provided for reference / portability to other runtimes.
 
 ## Model details
 
 - **Architecture:** LFM2 (`Lfm2ForCausalLM`) — hybrid short-convolution + grouped-query attention.
 - **Hidden size:** 1024 · **Layers:** 14 · **Attention heads:** 16 (8 KV heads, GQA) · **Intermediate size:** 2560.
-- **Vocabulary:** 65,536 · **Context length:** up to 128 k.
+- **Vocabulary:** 65 536 · **Context length:** up to 128 k.
 - **Precision:** bf16 on the NPU.
 - **Target:** Synaptics SL2619, compiled with the Torq compiler.
 
@@ -44,8 +57,9 @@ reference / portability to other runtimes.
 
 Runs on the Synaptics Torq runtime via
 [synaptics-torq/torq-examples](https://github.com/synaptics-torq/torq-examples). Place
-`model.vmfb`, `token_embeddings.npy`, `config.json`, and `tokenizer.json` in a model
-directory and invoke the Torq LLM runner with `model.vmfb`.
+the model files in a directory and invoke the Torq LLM runner with either `model.vmfb`
+(monolithic) or `body.vmfb` + `lm_head.vmfb` (split, lower TTFT), alongside
+`token_embeddings.npy`, `config.json`, and `tokenizer.json`.
 
 ## License