Publish RWKV7-1.5B-20260805

README.md

@@ -81,7 +81,7 @@ optional TileLang inference implementation are distributed with this release.
 
 | Field | Value |
 | --- | --- |
-| Repository | `BlinkDL/RWKV7-1.5B-20260805` |
+| Repository | `aabbdev/RWKV7-1.5B-20260805` |
 | Architecture class | `Rwkv7ForCausalLM` |
 | Public size label | `1.5`B |
 | Source parameters | `1,527,668,736` |
@@ -101,6 +101,12 @@ optional TileLang inference implementation are distributed with this release.
 
 ## Transformers quickstart
 
+Install the supported runtime before loading remote code:
+
+```bash
+python -m pip install "transformers>=5.3,<6" "huggingface-hub>=1.5,<2"
+```
+
 The repository includes `configuration_rwkv7.py` and `modeling_rwkv7.py`, adapted
 from the Transformers RWKV-7 integration at commit
 [`4ad9ed0`](https://github.com/huggingface/transformers/commit/4ad9ed0747ed6ba75c787e8f9040dcd64b166ee2).
@@ -116,7 +122,7 @@ from transformers import (
     PreTrainedConfig,
 )
 
-model_id = "BlinkDL/RWKV7-1.5B-20260805"
+model_id = "aabbdev/RWKV7-1.5B-20260805"
 tokenizer = AutoTokenizer.from_pretrained(
     model_id,
     config=PreTrainedConfig(),
@@ -153,7 +159,7 @@ def assistant_content(completion, thinking, *, close_incomplete=False):
         return f"{reply.rstrip()}\n</think>".strip()
     return "" if thinking_block is None else reply[thinking_block.end():].strip()
 
-model_id = "BlinkDL/RWKV7-1.5B-20260805"
+model_id = "aabbdev/RWKV7-1.5B-20260805"
 tokenizer = AutoTokenizer.from_pretrained(
     model_id,
     config=PreTrainedConfig(),
@@ -257,14 +263,14 @@ Install the versions listed in `inference/requirements.txt`, then run the bundle
 interactive chat:
 
 ```bash
-python inference/generate.py --model BlinkDL/RWKV7-1.5B-20260805 --backend auto --interactive
+python inference/generate.py --model aabbdev/RWKV7-1.5B-20260805 --backend auto --interactive
 ```
 
 Or independent prompts separated by blank lines:
 
 ```bash
 python inference/generate.py \
-  --model BlinkDL/RWKV7-1.5B-20260805 \
+  --model aabbdev/RWKV7-1.5B-20260805 \
   --backend auto \
   --input-file prompts.txt
 ```

configuration_rwkv7.py

@@ -17,6 +17,22 @@ from huggingface_hub.dataclasses import strict
 
 from transformers.configuration_utils import PreTrainedConfig
 from transformers.utils import auto_docstring
+from dataclasses import is_dataclass
+
+
+def _strict_config(cls):
+    # Transformers 5.15 makes PreTrainedConfig a dataclass; 5.3 does not. Hub's
+    # strict decorator is valid only in the former case. Rwkv7Config still runs its
+    # explicit __init__ validation on both branches.
+    if is_dataclass(cls):
+        return strict(cls)
+
+    def _legacy_init(self, **kwargs):
+        PreTrainedConfig.__init__(self, **kwargs)
+        self.__post_init__()
+
+    cls.__init__ = _legacy_init
+    return cls
 
 
 @auto_docstring(
@@ -29,7 +45,7 @@ from transformers.utils import auto_docstring
     (`BlinkDL/RWKV-LM`) rather than a renamed variant.
     """,
 )
-@strict
+@_strict_config
 class Rwkv7Config(PreTrainedConfig):
     r"""
     vocab_size (`int`, *optional*, defaults to 65536):
@@ -147,7 +163,9 @@ class Rwkv7Config(PreTrainedConfig):
                 f"num_heads must be hidden_size // head_dim = {self.hidden_size // self.head_dim}, "
                 f"got {self.num_heads}"
             )
-        super().__post_init__(**kwargs)
+        parent_post_init = getattr(super(), "__post_init__", None)
+        if parent_post_init is not None:
+            parent_post_init(**kwargs)
 
 
 __all__ = ["Rwkv7Config"]

inference/requirements.txt

@@ -1,5 +1,5 @@
-transformers>=5.15,<6
-huggingface-hub>=0.34
+transformers>=5.3,<6
+huggingface-hub>=1.5,<2
 safetensors>=0.5
 jinja2>=3.1,<4
 tilelang==0.1.12

modeling_rwkv7.py

@@ -20,7 +20,135 @@ import torch
 from torch import nn
 
 from transformers import initialization as init
-from transformers.cache_utils import Cache, LinearAttentionLayer
+try:
+    from transformers.cache_utils import Cache, LinearAttentionLayer
+except ImportError:
+    from transformers.cache_utils import Cache
+
+    class LinearAttentionLayer:
+        """Backport of the recurrent-only cache surface introduced after Transformers 5.3."""
+
+        is_compileable = True
+        supports_early_init = False
+        is_sliding = False
+
+        def __init__(self, number_of_states: int = 1, **kwargs):
+            del kwargs
+            self.number_of_states = number_of_states
+            self.conv_states = dict.fromkeys(range(number_of_states))
+            self.recurrent_states = dict.fromkeys(range(number_of_states))
+            self.is_conv_states_initialized = dict.fromkeys(range(number_of_states), False)
+            self.is_recurrent_states_initialized = dict.fromkeys(range(number_of_states), False)
+            self.has_previous_state = dict.fromkeys(range(number_of_states), False)
+            self.conv_kernel_size = dict.fromkeys(range(number_of_states))
+            self.device = None
+            self.dtype = None
+            self.record_past = False
+
+        def __repr__(self):
+            return self.__class__.__name__
+
+        @property
+        def is_initialized(self):
+            return all(self.is_recurrent_states_initialized.values())
+
+        @property
+        def max_batch_size(self):
+            for state in self.recurrent_states.values():
+                if state is not None:
+                    return state.shape[0]
+            return 0
+
+        @property
+        def max_cache_len(self):
+            return -1
+
+        def lazy_initialization(
+            self,
+            conv_states=None,
+            recurrent_states=None,
+            state_idx: int = 0,
+            conv_kernel_size=None,
+        ):
+            if conv_states is not None:
+                if self.device is None:
+                    self.dtype, self.device = conv_states.dtype, conv_states.device
+                size = conv_states.shape[-1] if conv_kernel_size is None else conv_kernel_size
+                self.conv_kernel_size[state_idx] = size
+                self.conv_states[state_idx] = torch.zeros(
+                    (*conv_states.shape[:-1], size),
+                    dtype=conv_states.dtype,
+                    device=conv_states.device,
+                )
+                self.is_conv_states_initialized[state_idx] = True
+            if recurrent_states is not None:
+                if self.device is None:
+                    self.dtype, self.device = recurrent_states.dtype, recurrent_states.device
+                self.recurrent_states[state_idx] = torch.zeros_like(recurrent_states)
+                self.is_recurrent_states_initialized[state_idx] = True
+
+        def update_recurrent_state(self, recurrent_states, state_idx: int = 0, **kwargs):
+            del kwargs
+            if not self.is_recurrent_states_initialized[state_idx]:
+                self.lazy_initialization(recurrent_states=recurrent_states, state_idx=state_idx)
+            self.recurrent_states[state_idx].copy_(recurrent_states)
+            self.has_previous_state[state_idx] = True
+            return self.recurrent_states[state_idx]
+
+        def update(self, key_states, value_states, cache_kwargs=None):
+            del key_states, value_states, cache_kwargs
+            raise NotImplementedError("RWKV7 updates recurrent state through update_recurrent_state")
+
+        def get_seq_length(self):
+            return 0
+
+        def get_mask_sizes(self, cache_position):
+            return cache_position.shape[0], 0
+
+        def get_max_cache_shape(self):
+            return -1
+
+        def reset(self):
+            for state_idx in range(self.number_of_states):
+                state = self.recurrent_states[state_idx]
+                if state is not None:
+                    state.zero_()
+                self.has_previous_state[state_idx] = False
+
+        def reorder_cache(self, beam_idx):
+            for state_idx in range(self.number_of_states):
+                state = self.recurrent_states[state_idx]
+                if state is not None:
+                    self.recurrent_states[state_idx] = state.index_select(0, beam_idx.to(state.device))
+
+        def batch_repeat_interleave(self, repeats: int):
+            for state_idx in range(self.number_of_states):
+                state = self.recurrent_states[state_idx]
+                if state is not None:
+                    self.recurrent_states[state_idx] = state.repeat_interleave(repeats, dim=0)
+
+        def batch_select_indices(self, indices):
+            for state_idx in range(self.number_of_states):
+                state = self.recurrent_states[state_idx]
+                if state is not None:
+                    self.recurrent_states[state_idx] = state[indices, ...]
+
+        def offload(self):
+            for state_idx in range(self.number_of_states):
+                state = self.recurrent_states[state_idx]
+                if state is not None:
+                    self.recurrent_states[state_idx] = state.to("cpu", non_blocking=True)
+
+        def prefetch(self):
+            for state_idx in range(self.number_of_states):
+                state = self.recurrent_states[state_idx]
+                if state is not None and self.device is not None and state.device != self.device:
+                    self.recurrent_states[state_idx] = state.to(self.device, non_blocking=True)
+
+        def crop(self, max_length):
+            if max_length not in (0, -1):
+                raise RuntimeError("RWKV7 recurrent cache does not support rollback")
+
 from transformers.generation import GenerationMixin
 from transformers.modeling_layers import GradientCheckpointingLayer
 from transformers.modeling_utils import PreTrainedModel
@@ -749,8 +877,10 @@ class Rwkv7Cache(Cache):
                 layer.is_recurrent_states_initialized[slot] = True
                 layer.has_previous_state[slot] = True
             return
+        layer = self.layers[layer_idx]
+        assert isinstance(layer, Rwkv7CacheLayer)
         for slot, state in states:
-            self.update_recurrent_state(state, layer_idx, slot)
+            layer.update_recurrent_state(state, slot)
 
 
 class Rwkv7Block(GradientCheckpointingLayer):
@@ -834,6 +964,7 @@ class Rwkv7CausalLMOutput(ModelOutput):
     loss: torch.FloatTensor | None = None
     logits: torch.FloatTensor | None = None
     state: Rwkv7Cache | None = None
+    past_key_values: Rwkv7Cache | None = None
     hidden_states: tuple[torch.FloatTensor, ...] | None = None
     attentions: None = None
 
@@ -1119,6 +1250,26 @@ class Rwkv7ForCausalLM(Rwkv7PreTrainedModel, GenerationMixin):
     def set_output_embeddings(self, new_embeddings):
         self.head = new_embeddings
 
+    @staticmethod
+    def _expand_inputs_for_generation(
+        expand_size=1,
+        is_encoder_decoder=False,
+        input_ids=None,
+        **model_kwargs,
+    ):
+        cache = model_kwargs.get("state")
+        if cache is None:
+            cache = model_kwargs.get("past_key_values")
+        input_ids, model_kwargs = GenerationMixin._expand_inputs_for_generation(
+            expand_size=expand_size,
+            is_encoder_decoder=is_encoder_decoder,
+            input_ids=input_ids,
+            **model_kwargs,
+        )
+        if cache is not None and expand_size > 1:
+            cache.batch_repeat_interleave(expand_size)
+        return input_ids, model_kwargs
+
     def prepare_inputs_for_generation(
         self,
         input_ids,
@@ -1128,6 +1279,9 @@ class Rwkv7ForCausalLM(Rwkv7PreTrainedModel, GenerationMixin):
         is_first_iteration=False,
         **kwargs,
     ):
+        legacy_state = kwargs.pop("past_key_values", None)
+        if state is None:
+            state = legacy_state
         # `state is not None` does not by itself mean decode: callers can provide
         # an empty preallocated state for the initial prompt, or a carried state
         # followed by a multi-token continuation. GenerationMixin tells us how
@@ -1165,6 +1319,7 @@ class Rwkv7ForCausalLM(Rwkv7PreTrainedModel, GenerationMixin):
         position_ids: torch.LongTensor | None = None,
         inputs_embeds: torch.FloatTensor | None = None,
         state: Rwkv7Cache | None = None,
+        past_key_values: Rwkv7Cache | None = None,
         labels: torch.LongTensor | None = None,
         use_cache: bool | None = None,
         output_attentions: bool | None = None,
@@ -1185,6 +1340,8 @@ class Rwkv7ForCausalLM(Rwkv7PreTrainedModel, GenerationMixin):
             so `generate` declined to pass it and a caller who passed it was quietly
             ignored.
         """
+        if state is None:
+            state = past_key_values
         if labels is not None and use_cache is None:
             use_cache = False
         outputs = self.rwkv7(
@@ -1209,7 +1366,13 @@ class Rwkv7ForCausalLM(Rwkv7PreTrainedModel, GenerationMixin):
         if labels is not None:
             loss = self.loss_function(logits, labels, self.config.vocab_size, **kwargs)
 
-        return Rwkv7CausalLMOutput(loss=loss, logits=logits, state=outputs.state, hidden_states=outputs.hidden_states)
+        return Rwkv7CausalLMOutput(
+            loss=loss,
+            logits=logits,
+            state=outputs.state,
+            past_key_values=outputs.state,
+            hidden_states=outputs.hidden_states,
+        )
 
 
 __all__ = ["Rwkv7Cache", "Rwkv7PreTrainedModel", "Rwkv7Model", "Rwkv7ForCausalLM"]

release-manifest.json

@@ -18,6 +18,7 @@
     "tensor_count": 798,
     "tensor_map_sha256": "03131ced241e7b0f86869363b3969ead6ef58462f23d9f04e56e26d00565aefc"
   },
+  "derivation": null,
   "files": {
     ".gitattributes": {
       "role": "metadata",
@@ -36,8 +37,8 @@
     },
     "README.md": {
       "role": "model_card",
-      "sha256": "fbcb25fd1bb80a23a961c9bc85ff57f606177341e7d24143a5cbcc75b69413da",
-      "size_bytes": 11454
+      "sha256": "62d2cfb899bd1ed68bef0c5b51fe26f88937c87f8dbe19482a99f1d35d3f9794",
+      "size_bytes": 11597
     },
     "chat_template.jinja": {
       "role": "tokenizer",
@@ -51,8 +52,8 @@
     },
     "configuration_rwkv7.py": {
       "role": "model_code",
-      "sha256": "a4a2adcd02a101cc2a17f2452f8e56f4a8d8348be0d19ab989e908a83921c123",
-      "size_bytes": 7318
+      "sha256": "2ef125dc431b540ada5fec14e0df16cc9af6c921ece55f13016c7d323ca5f98a",
+      "size_bytes": 7933
     },
     "generation_config.json": {
       "role": "model_config",
@@ -76,8 +77,8 @@
     },
     "inference/requirements.txt": {
       "role": "inference",
-      "sha256": "9cc4f89e2edc45da7a04cdf8e4c1da83b682a6bf969e0eed7b3dec91128d925a",
-      "size_bytes": 93
+      "sha256": "289e514cb8a38aef51615a254abb514d0eed4556f4901eb910a1191e9e084199",
+      "size_bytes": 94
     },
     "inference/runtime.py": {
       "role": "inference",
@@ -91,8 +92,8 @@
     },
     "modeling_rwkv7.py": {
       "role": "model_code",
-      "sha256": "8b8a3459b40a33424b592abff360485577a8f19419be411e3dceb817a51ebe46",
-      "size_bytes": 58685
+      "sha256": "7134c72d46ca6b6b8b152444d546bc2a129f8ea430279fa3baf6507a101800e5",
+      "size_bytes": 65301
     },
     "tokenizer.json": {
       "role": "tokenizer",
@@ -139,8 +140,9 @@
     "provenance": "locked-profile"
   },
   "model_code": {
-    "format_version": 2,
+    "format_version": 3,
     "patches": [
+      "transformers-5.3-config-and-cache-compatibility",
       "layer-zero-value-residual-buffers",
       "trainer-past-key-values-placeholder",
       "trl-position-ids-packing-boundaries",
@@ -151,18 +153,19 @@
     "sources": {
       "configuration_rwkv7.py": {
         "asset_path": "model_code/configuration_rwkv7.py",
-        "output_sha256": "a4a2adcd02a101cc2a17f2452f8e56f4a8d8348be0d19ab989e908a83921c123",
+        "output_sha256": "2ef125dc431b540ada5fec14e0df16cc9af6c921ece55f13016c7d323ca5f98a",
         "repository_path": "src/transformers/models/rwkv7/configuration_rwkv7.py",
         "source_sha256": "6f5b92c5fe7498ad22b0054a2f735a7ca82e7577436f4ad32f0fc27d1e900fdd"
       },
       "modeling_rwkv7.py": {
         "asset_path": "model_code/modeling_rwkv7.py",
-        "output_sha256": "8b8a3459b40a33424b592abff360485577a8f19419be411e3dceb817a51ebe46",
+        "output_sha256": "7134c72d46ca6b6b8b152444d546bc2a129f8ea430279fa3baf6507a101800e5",
         "repository_path": "src/transformers/models/rwkv7/modeling_rwkv7.py",
         "source_sha256": "3e8e5af7c4eba0b5de1496aef44773d7ac1bb4d96756e6f55efaf29453d67952"
       }
     },
-    "transformers_min_version": "5.15"
+    "transformers_max_version": "6",
+    "transformers_min_version": "5.3"
   },
   "profile": {
     "checkpoint": "g1i-1.5b-20260805",
@@ -326,7 +329,7 @@
       }
     }
   },
-  "schema_version": 8,
+  "schema_version": 9,
   "source": {
     "filename": "rwkv7-g1i-1.5b-20260805-ctx16384.pth",
     "kind": "huggingface",