Use chunked WKV by default

Apply rwkv-publisher upstream commit 0e23b39775d6017638d0234d153bd263c411c229. The existing chunked prefill path becomes the default; weights remain unchanged.

README.md

@@ -132,6 +132,12 @@ model = AutoModelForCausalLM.from_pretrained(
 The recurrent cache returned by the model can be passed back for incremental
 decoding. Use an `attention_mask` for padded batches.
 
+The model defaults to the chunk-parallel WKV path for efficient multi-token
+prefill. To reproduce the portable token-order reference path, set
+`model.config.wkv_implementation = "eager"` before the first forward pass.
+Chunked execution changes floating-point operation order, so small numerical
+differences from eager execution are expected.
+
 ## Chat quickstart
 
 ```python

config.json

@@ -26,6 +26,6 @@
   "use_cache": true,
   "v_low_rank_dim": 64,
   "vocab_size": 65536,
-  "wkv_implementation": "eager",
+  "wkv_implementation": "chunked",
   "wkv_state_dtype": "float32"
 }

configuration_rwkv7.py

@@ -70,7 +70,7 @@ class Rwkv7Config(PreTrainedConfig):
         the whole sequence, so a narrow state drifts; `"float32"` with fp16
         activations is the combination the reference implementation uses.
         `"float16"`/`"bfloat16"` trade that for a smaller state.
-    wkv_implementation (`str`, *optional*, defaults to `"eager"`):
+    wkv_implementation (`str`, *optional*, defaults to `"chunked"`):
         Which WKV recurrence to use, by name, from
         `models.rwkv7.modeling_rwkv7.RWKV7_WKV_FUNCTIONS`. `"eager"` is the
         exact portable PyTorch path and preserves the reference token order for
@@ -120,7 +120,7 @@ class Rwkv7Config(PreTrainedConfig):
     tie_word_embeddings: bool = False
     use_cache: bool = True
     wkv_state_dtype: str = "float32"
-    wkv_implementation: str = "eager"
+    wkv_implementation: str = "chunked"
     bos_token_id: int | None = 0
     eos_token_id: int | None = 0
     pad_token_id: int | None = 0

release-manifest.json

@@ -36,8 +36,8 @@
     },
     "README.md": {
       "role": "model_card",
-      "sha256": "510b13234261aea6adf15b475db2d901b5af3087878221d4f200620a935d086c",
-      "size_bytes": 11815
+      "sha256": "9e71f101f9d45fb457d197117bb0df15b8d6cb65e2acbbdf3cfa25c54141e36d",
+      "size_bytes": 12164
     },
     "assets/rwkv-logo.webp": {
       "role": "media",
@@ -51,13 +51,13 @@
     },
     "config.json": {
       "role": "model_config",
-      "sha256": "4bf76c77a139beb6a5a949cf4815724e20e4ec380c78670b7aaca6648b018f38",
-      "size_bytes": 779
+      "sha256": "95f3524181116fca466ce5b2e8c827dcf1a026ebfb4967d3ad28af9feee96d1f",
+      "size_bytes": 781
     },
     "configuration_rwkv7.py": {
       "role": "model_code",
-      "sha256": "95b6a1afabaffe7c645a89471779f773cb2bce9326c2f62630f2d2a00133305f",
-      "size_bytes": 7427
+      "sha256": "14050c631e27d66cd2f52ca5ec53b62eef04e602c4f1fc4e3696c328bc8e200e",
+      "size_bytes": 7431
     },
     "generation_config.json": {
       "role": "model_config",
@@ -169,37 +169,38 @@
     "provenance": "locked-profile"
   },
   "model_code": {
-    "format_version": 4,
+    "format_version": 5,
+    "source_repository": "https://github.com/huggingface/transformers.git",
+    "source_revision": "4ad9ed0747ed6ba75c787e8f9040dcd64b166ee2",
+    "transformers_min_version": "5.15",
     "patches": [
+      "chunked-wkv-default",
       "model-neutral-configuration-docstring",
       "layer-zero-value-residual-buffers",
       "trainer-past-key-values-placeholder",
       "trl-position-ids-packing-boundaries",
       "labels-disable-recurrent-cache"
     ],
-    "source_repository": "https://github.com/huggingface/transformers.git",
-    "source_revision": "4ad9ed0747ed6ba75c787e8f9040dcd64b166ee2",
     "sources": {
       "configuration_rwkv7.py": {
         "asset_path": "model_code/configuration_rwkv7.py",
-        "output_sha256": "ad5db5a9d335015e316ae5a4ff3913c48b592424513881c4379f2ad732116fd9",
         "repository_path": "src/transformers/models/rwkv7/configuration_rwkv7.py",
-        "source_sha256": "6f5b92c5fe7498ad22b0054a2f735a7ca82e7577436f4ad32f0fc27d1e900fdd"
+        "source_sha256": "6f5b92c5fe7498ad22b0054a2f735a7ca82e7577436f4ad32f0fc27d1e900fdd",
+        "output_sha256": "76f832fa713fd7e117273a67a86225fe2faebc6d80a6757113484def3a5671b4"
       },
       "modeling_rwkv7.py": {
         "asset_path": "model_code/modeling_rwkv7.py",
-        "output_sha256": "8b8a3459b40a33424b592abff360485577a8f19419be411e3dceb817a51ebe46",
         "repository_path": "src/transformers/models/rwkv7/modeling_rwkv7.py",
-        "source_sha256": "3e8e5af7c4eba0b5de1496aef44773d7ac1bb4d96756e6f55efaf29453d67952"
+        "source_sha256": "3e8e5af7c4eba0b5de1496aef44773d7ac1bb4d96756e6f55efaf29453d67952",
+        "output_sha256": "8b8a3459b40a33424b592abff360485577a8f19419be411e3dceb817a51ebe46"
       },
       "tokenization_rwkv7.py": {
         "asset_path": "model_code/tokenization_rwkv7.py",
-        "output_sha256": "20c29d0f8f27003889570f54a1cd53482dd608a034e87673f6e43cf5e7c8d173",
         "repository_path": "rwkv-publisher/src/rwkv_publisher/assets/model_code/tokenization_rwkv7.py",
-        "source_sha256": "20c29d0f8f27003889570f54a1cd53482dd608a034e87673f6e43cf5e7c8d173"
+        "source_sha256": "20c29d0f8f27003889570f54a1cd53482dd608a034e87673f6e43cf5e7c8d173",
+        "output_sha256": "20c29d0f8f27003889570f54a1cd53482dd608a034e87673f6e43cf5e7c8d173"
       }
-    },
-    "transformers_min_version": "5.15"
+    }
   },
   "profile": {
     "checkpoint": "g1j-1.5b-20260831",
@@ -372,4 +373,4 @@
     "sha256": "c43176881caf85fe22ad654ab02e7519260d560f3d20420ab590adb0c823860f",
     "size_bytes": 3055444605
   }
-}
+}