Chat template: accept the 0.18 content-parts form (string form unchanged); weights, tokenizer and executor metadata byte-identical

README.md

@@ -63,6 +63,8 @@ Two ways to read the answer:
 | `Shieldstral-1.0-3B-vision_int4.litertlm` | **text + image** | int4-b32 decoder + int8 pixtral tower, static 560×560 | 2.78 GB | 1.82 GiB | **image moderation, phones and up** |
 | `Shieldstral-1.0-3B-vision_int4_gpu.litertlm` | **text + image** | the same weights and the same tower; decoder re-exported so the attention softmax lowers to a builtin | 2.78 GB | 1.82 GiB | **the vision file for GPU** |
 
+2026-09-21: chat template updated to accept the 0.18 content-parts form (string form unchanged); weights, tokenizer and executor metadata byte-identical.
+
 **Pick int4 unless you have a reason not to.** On the gate set it matches int8 on every metric, and it is the only text variant that fits an iPhone: int8's 3.33 GiB single section exceeds the practical iOS mmap budget (~2.1 GiB for an app with default entitlements; even entitlement-relaxed apps have topped out below 3 GiB on current hardware). All variants share an identical int8 embedding section, so the difference is in the decoder weights.
 
 The vision bundle accepts text documents too, so it can replace the text one — it just costs 0.6 GB more on disk and loads the tower you may not use. **On GPU, use `_int4_gpu` for that.** The original vision file does not load on a mobile GPU at all: litert-torch 0.9.2 marks the attention softmax as an `odml.softmax` StableHLO composite and litert-converter 0.3.0 cannot lower it, so the delegate takes 52 of 1187 ops and the engine is refused. The text files were exported through a path that stripped the marker, which is why they were unaffected. `_int4_gpu` is the same weights and the same pixtral tower, re-exported on litert-converter 0.3.1, which lowers the composite to a builtin `SOFTMAX`.

Shieldstral-1.0-3B-vision_int4.litertlm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:180b948a4d1d4eacb5219c1124429ee9be392367dcb558e255f7ed466f0f4092
+oid sha256:3f9289463889fe1232da2bee6ab133804b3e5a381808b00dc1b29e9e36b877bc
 size 2783331824

Shieldstral-1.0-3B-vision_int4_gpu.litertlm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:487639d5e90ff600dd3a9515a9526f11d4f6959c58c2db9bd411c94c1bf6d4d7
+oid sha256:9dcd46ff64ba528bf57e6153a97c068364efc48c14f36eeb7cada9a56dea7a5b
 size 2783086064

litertlm_manifest.json

@@ -1,7 +1,7 @@
 {
   "manifest_schema": "0.1.2",
   "repo": "litert-community/Shieldstral-1.0-3B",
-  "generated": "2026-09-06",
+  "generated": "2026-09-21",
   "generator": "make_manifest.py",
   "model": {
     "display_name": "Shieldstral-1.0-3B",
@@ -21,12 +21,12 @@
   "variants": [
     {
       "file": "Shieldstral-1.0-3B-vision_int4.litertlm",
-      "sha256": "180b948a4d1d4eacb5219c1124429ee9be392367dcb558e255f7ed466f0f4092",
+      "sha256": "3f9289463889fe1232da2bee6ab133804b3e5a381808b00dc1b29e9e36b877bc",
       "size_bytes": 2783331824,
       "sections": [
         {
           "type": "LlmMetadataProto",
-          "size_bytes": 892
+          "size_bytes": 1175
         },
         {
           "type": "HF_Tokenizer_Zlib",
@@ -82,12 +82,12 @@
     },
     {
       "file": "Shieldstral-1.0-3B-vision_int4_gpu.litertlm",
-      "sha256": "487639d5e90ff600dd3a9515a9526f11d4f6959c58c2db9bd411c94c1bf6d4d7",
+      "sha256": "9dcd46ff64ba528bf57e6153a97c068364efc48c14f36eeb7cada9a56dea7a5b",
       "size_bytes": 2783086064,
       "sections": [
         {
           "type": "LlmMetadataProto",
-          "size_bytes": 892
+          "size_bytes": 1175
         },
         {
           "type": "HF_Tokenizer_Zlib",