diff --git a/cli/alora/train.py b/cli/alora/train.py
index 722a20096..65581c0be 100644
--- a/cli/alora/train.py
+++ b/cli/alora/train.py
@@ -20,6 +20,7 @@
     from transformers import (
         AutoModelForCausalLM,
         AutoTokenizer,
+        PreTrainedTokenizerBase,
         TrainerCallback,
         TrainerControl,
         TrainerState,
@@ -53,7 +54,7 @@
 
 
 def load_dataset_from_json(
-    json_path: str, tokenizer: AutoTokenizer, invocation_prompt: str
+    json_path: str, tokenizer: PreTrainedTokenizerBase, invocation_prompt: str
 ) -> Dataset:
     """Load a JSONL dataset and format it for SFT training.
 
@@ -165,8 +166,12 @@ def save_model(self, output_dir: str | None = None, _internal_call: bool = False
         """
         if self.model is not None:
             self.model.save_pretrained(output_dir, safe_serialization=True)
-            if self.tokenizer is not None:
-                self.tokenizer.save_pretrained(output_dir)
+            # transformers v5 renamed .tokenizer -> .processing_class
+            processor = getattr(self, "processing_class", None) or getattr(
+                self, "tokenizer", None
+            )
+            if processor is not None:
+                processor.save_pretrained(output_dir)
 
 
 def train_model(
@@ -224,7 +229,8 @@ def train_model(
         base_model, padding_side="right", trust_remote_code=True
     )
     tokenizer.pad_token = tokenizer.eos_token
-    tokenizer.add_special_tokens = False
+    # Note: previously had `tokenizer.add_special_tokens = False` here, which was
+    # a no-op (assigning False to a method reference). Removed — do not restore.
 
     dataset = load_dataset_from_json(dataset_path, tokenizer, invocation_prompt)
     dataset = dataset.shuffle(seed=42)
diff --git a/docs/docs/advanced/prefix-caching-and-kv-blocks.md b/docs/docs/advanced/prefix-caching-and-kv-blocks.md
index a69e53e44..751001e08 100644
--- a/docs/docs/advanced/prefix-caching-and-kv-blocks.md
+++ b/docs/docs/advanced/prefix-caching-and-kv-blocks.md
@@ -64,7 +64,7 @@ When a prompt contains a mix of cached and uncached blocks, Mellea:
 2. Runs forward passes on uncached blocks.
 3. Retrieves stored `DynamicCache` for cached blocks.
 4. **Smashes** (concatenates) all KV caches along the time axis using
-   `merge_dynamic_caches()`.
+   `merge_dynamic_caches_v5()`.
 5. Passes the merged cache plus the combined input IDs to the generation step.
 
 The result is identical to a single full-context forward pass, with the prefill
diff --git a/docs/kv_smash/kv_with_chat.py b/docs/kv_smash/kv_with_chat.py
index 7f43c5398..3e3d91527 100644
--- a/docs/kv_smash/kv_with_chat.py
+++ b/docs/kv_smash/kv_with_chat.py
@@ -1,7 +1,7 @@
 import torch
 
 from mellea.backends.huggingface import LocalHFBackend
-from mellea.backends.kv_block_helpers import DynamicCache, merge_dynamic_caches
+from mellea.backends.kv_block_helpers import DynamicCache, merge_dynamic_caches_v5
 from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 
 backend = LocalHFBackend(model_id=IBM_GRANITE_4_HYBRID_MICRO)
@@ -30,7 +30,7 @@ def cache(s: str, store=True) -> DynamicCache:
 def merge(toks, dcs):
     merged_toks = torch.cat([t["input_ids"] for t in toks], dim=1)
     merged_masks = torch.cat([t["attention_mask"] for t in toks], dim=1)
-    merged_dcs = merge_dynamic_caches(dcs)
+    merged_dcs = merge_dynamic_caches_v5(dcs)
 
     return merged_toks, merged_masks, merged_dcs
 
@@ -89,7 +89,7 @@ def merge(toks, dcs):
 # Merge everything together.
 merged_toks = torch.cat([toks["input_ids"] for toks in tok_parts], dim=1)
 merged_masks = torch.cat([toks["attention_mask"] for toks in tok_parts], dim=1)
-merged_dcs = merge_dynamic_caches(dc_parts)
+merged_dcs = merge_dynamic_caches_v5(dc_parts)
 
 # crop the last KV for safety.
 merged_dcs.crop(-1)
diff --git a/docs/kv_smash/kvcache.py b/docs/kv_smash/kvcache.py
index 83f226cf1..850bff7b5 100644
--- a/docs/kv_smash/kvcache.py
+++ b/docs/kv_smash/kvcache.py
@@ -4,17 +4,20 @@
 #   "mellea[hf]",
 # ]
 # ///
+from typing import cast
+
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizer
-from transformers.generation import GenerateDecoderOnlyOutput
+from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
+from transformers.generation.utils import GenerateDecoderOnlyOutput
+from transformers.modeling_utils import PreTrainedModel
 
-from mellea.backends.kv_block_helpers import DynamicCache, merge_dynamic_caches
+from mellea.backends.kv_block_helpers import DynamicCache, merge_dynamic_caches_v5
 
 model_id = "ibm-granite/granite-4.0-tiny-preview"
 device = torch.device("mps")
-model = AutoModelForCausalLM.from_pretrained(model_id)
+model: PreTrainedModel = AutoModelForCausalLM.from_pretrained(model_id)  # type: ignore[assignment]
 # model = model.to(device=device) # this part does not pass mypy; possible misconfiguration
-tokenizer: PreTrainedTokenizer = AutoTokenizer.from_pretrained(model_id)
+tokenizer: PreTrainedTokenizerBase = AutoTokenizer.from_pretrained(model_id)
 
 
 def cache(toks) -> DynamicCache:
@@ -34,7 +37,7 @@ def merge(strs: list[str]):
 
     merged_toks = torch.cat([toks["input_ids"] for toks in strs_toks], dim=1)
     merged_masks = torch.cat([toks["attention_mask"] for toks in strs_toks], dim=1)
-    merged_dcs = merge_dynamic_caches(strs_dcs)
+    merged_dcs = merge_dynamic_caches_v5(strs_dcs)
 
     return merged_toks, merged_masks, merged_dcs
 
@@ -45,7 +48,7 @@ def merge(strs: list[str]):
 merged_dcs.crop(-1)
 
 # GenerateDecoderOnlyOutput | GenerateEncoderDecoderOutput | GenerateBeamDecoderOnlyOutput | GenerateBeamEncoderDecoderOutput | LongTensor
-result = model.generate(
+result = model.generate(  # type: ignore[operator]
     merged_toks.to(model.device),
     attention_mask=merged_masks.to(model.device),
     past_key_values=merged_dcs,
diff --git a/docs/metrics/coverage-current.json b/docs/metrics/coverage-current.json
index 80d7493e9..9c04e28a7 100644
--- a/docs/metrics/coverage-current.json
+++ b/docs/metrics/coverage-current.json
@@ -355,7 +355,7 @@
       "TokenizedCacheIterleaving",
       "LegacyCache",
       "legacy_cache_smash",
-      "merge_dynamic_caches",
+      "merge_dynamic_caches_v5",
       "tokens_to_legacy_cache"
     ],
     "mellea.backends.huggingface.granite_formatters": [
diff --git a/mellea/backends/huggingface.py b/mellea/backends/huggingface.py
index 4f14c36c3..ea28e1a77 100644
--- a/mellea/backends/huggingface.py
+++ b/mellea/backends/huggingface.py
@@ -12,7 +12,7 @@
 import json
 import threading
 from collections.abc import Callable, Coroutine, Sequence
-from typing import Any, overload
+from typing import Any, cast, overload
 
 import jinja2
 import jinja2.meta
@@ -32,7 +32,7 @@
     from transformers.generation.streamers import AsyncTextIteratorStreamer
     from transformers.generation.utils import GenerateDecoderOnlyOutput
     from transformers.modeling_utils import PreTrainedModel
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers.tokenization_utils_base import PreTrainedTokenizerBase
     from transformers.trainer_utils import set_seed
 except ImportError as e:
     raise ImportError(
@@ -125,7 +125,7 @@ def _install_cancel_stopping_criteria(
 
 Huggingface backends can initialize themselves from a model string if the transformers `Auto*` classes can be used. Therefore, a TransformersTorchConfig usually isn't required. However, sometimes a model needs special care to instantiate properly, or a custom device type needs to bse used. Instead of trying to do a lot of partial magic, we basically have two modaliites: either the constructor can figure out everything from the model_id, or the user has to provide an entire config.
 """
-TransformersTorchConfig = tuple[PreTrainedTokenizer, PreTrainedModel, torch.device]
+TransformersTorchConfig = tuple[PreTrainedTokenizerBase, PreTrainedModel, torch.device]
 
 format: None = None  # typing this variable in order to shadow the global format function and ensure mypy checks for errors
 
@@ -213,16 +213,21 @@ def _cleanup_kv_cache(cache_info: HFAloraCacheInfo) -> None:
 # group) or silently replace a function injected by the Jinja environment (for the globals group).
 #
 # ⚠️  REVIEW NOTE: verify this set against the transformers source whenever upgrading.
-# Named params: transformers.tokenization_utils_base.PreTrainedTokenizerBase.apply_chat_template
-#               (the render_jinja_template call around line 119 in that method)
-# Jinja globals: transformers.utils.chat_template_utils._compile_jinja_template
-#               (the jinja_env.globals assignments near the bottom of that function)
+# Namespace vars: transformers.utils.chat_template_utils.render_jinja_template
+#                 (the compiled_template.render(...) call — every kwarg there becomes
+#                 a Jinja variable and must be excluded from the allowlist)
+# Jinja globals: transformers.utils.chat_template_utils._cached_compile_jinja_template
+#                (the jinja_env.globals[...] assignments near the bottom)
 _HF_INTERNAL_TEMPLATE_VARS: frozenset[str] = frozenset(
     {
-        # --- Named parameters passed explicitly to render_jinja_template ---
-        # Forwarding these from model_options would cause duplicate-kwarg TypeError.
-        "messages",  # the conversation; always the first positional arg
+        # --- Variables injected into the Jinja namespace by render_jinja_template ---
+        # These come from `compiled_template.render(messages=chat, tools=..., documents=...,
+        # add_generation_prompt=..., **kwargs)` in transformers' chat_template_utils.
+        # Forwarding any of these from model_options to apply_chat_template would
+        # cause a duplicate-kwarg TypeError or silently shadow the value HF supplies.
+        "messages",  # the conversation; HF binds `chat` to this name in render()
         "tools",  # tool schemas; our call sites pass tools=convert_tools_to_json(...) explicitly
+        "documents",  # RAG documents; injected even when None
         "add_generation_prompt",  # bool; passed explicitly at the standard-generation call site
         # --- Jinja environment globals set by _compile_jinja_template ---
         # find_undeclared_variables cannot see env globals, so these appear as "undeclared"
@@ -328,8 +333,8 @@ def __init__(
                 self._model: PreTrainedModel = AutoModelForCausalLM.from_pretrained(
                     self._hf_model_id, device_map=str(self._device), torch_dtype="auto"
                 )
-                self._tokenizer: PreTrainedTokenizer = AutoTokenizer.from_pretrained(
-                    self._hf_model_id
+                self._tokenizer: PreTrainedTokenizerBase = (
+                    AutoTokenizer.from_pretrained(self._hf_model_id)
                 )
             case _:
                 self._tokenizer, self._model, self._device = custom_config
@@ -844,7 +849,7 @@ def _make_merged_kv_cache(
             [toks["attention_mask"] for toks in tok_parts], dim=1
         )
         assert input_ids.shape == attention_mask.shape
-        merged_cache: DynamicCache = kv_block_helpers.merge_dynamic_caches(dc_parts)
+        merged_cache: DynamicCache = kv_block_helpers.merge_dynamic_caches_v5(dc_parts)
         # TODO: also assert that the merged cached is the correct shape given the input_ids and attention_mask shapes.
         # rewind merged cache by 1 for safety.
         merged_cache.crop(-1)  # type: ignore
@@ -1127,7 +1132,8 @@ async def _generate_from_context_standard(
                 "",  # Empty for no adapters.
                 self._model.generate,  # type: ignore
                 # Passed as args/kwargs to generate.
-                input_ids,
+                inputs=input_ids["input_ids"],
+                attention_mask=input_ids["attention_mask"],
                 return_dict_in_generate=True,
                 use_cache=self._use_caches,  # Only create KV cache if caching is enabled
                 **generate_kwargs,
@@ -1207,6 +1213,10 @@ async def processing(
             input_ids: The prompt token IDs used for decoding; required to slice off
                 the prompt portion from the generated sequences.
         """
+        input_ids_tensor = (
+            input_ids if isinstance(input_ids, torch.Tensor) else input_ids["input_ids"]
+        )
+
         if mot._underlying_value is None:
             mot._underlying_value = ""
 
@@ -1217,8 +1227,12 @@ async def processing(
         elif isinstance(chunk, GenerateDecoderOnlyOutput):
             # Otherwise, it's a non-streaming request. Decode it here.
             mot._meta["hf_output"] = chunk
-            mot._underlying_value += self._tokenizer.decode(
-                chunk.sequences[0, input_ids.shape[1] :], skip_special_tokens=True
+            mot._underlying_value += cast(
+                str,
+                self._tokenizer.decode(
+                    chunk.sequences[0, input_ids_tensor.shape[1] :],
+                    skip_special_tokens=True,
+                ),
             )
 
     async def post_processing(
@@ -1272,7 +1286,11 @@ class used during generation, if any.
                 kv_cache=kv_cache,
                 merged_token_ids=output_complete,
                 merged_attention=torch.ones_like(output_complete).to(self._device),
-                q_end=len(input_ids[0]),  # type: ignore
+                q_end=(
+                    input_ids
+                    if isinstance(input_ids, torch.Tensor)
+                    else input_ids["input_ids"]
+                ).shape[1],
                 scores=hf_output.scores,
             )
 
@@ -1302,7 +1320,11 @@ class used during generation, if any.
         if isinstance(hf_output, GenerateDecoderOnlyOutput):
             try:
                 if input_ids is not None and hf_output.sequences is not None:
-                    n_prompt = input_ids.shape[1]
+                    n_prompt = (
+                        input_ids
+                        if isinstance(input_ids, torch.Tensor)
+                        else input_ids["input_ids"]
+                    ).shape[1]
                     n_completion = hf_output.sequences[0].shape[0] - n_prompt
             except Exception:
                 pass
@@ -1772,14 +1794,11 @@ def load_adapter(self, adapter_qualified_name: str):
             )
 
         try:
-            adapter_kwargs = {}
-
-            # Peft tries to stringify the device. If it's mps, it gets stringified as "mps:0" which causes
-            # an error when loading with safetensors.torch.load_file. Force the device as a string "mps" to fix.
-            if self._device == torch.device("mps"):
-                adapter_kwargs["device"] = "mps"
+            # v5: adapter_kwargs is forwarded to download_kwargs only; device is
+            # derived automatically from self.device, so we don't pass it here —
+            # find_adapter_config_file() no longer accepts a 'device' argument.
             self._model.load_adapter(
-                adapter.path, adapter.qualified_name, adapter_kwargs=adapter_kwargs
+                adapter.path, adapter.qualified_name, adapter_kwargs={}
             )
         except ValueError as e:
             # If it's just that it's already loaded, ignore it.
diff --git a/mellea/backends/kv_block_helpers.py b/mellea/backends/kv_block_helpers.py
index b3a8e5dd9..8b35c8cab 100644
--- a/mellea/backends/kv_block_helpers.py
+++ b/mellea/backends/kv_block_helpers.py
@@ -1,20 +1,21 @@
 """Low-level utilities for concatenating transformer KV caches (KV smashing).
 
-Provides functions for merging `DynamicCache` and legacy tuple caches along the
-time axis (`merge_dynamic_caches`, `legacy_cache_smash`), and
-`tokens_to_legacy_cache` for converting a tokenized prompt into a prefilled KV
-cache. These helpers are used internally by local HuggingFace backends that reuse
-cached prefix computations across multiple generation calls.
+Provides ``prefill_cache_v5`` for converting a tokenized prompt into a prefilled
+``DynamicCache``, ``merge_dynamic_caches_v5`` for concatenating multiple
+``DynamicCache`` objects along the time axis, and ``merge_v5`` which composes
+the two. These helpers are used internally by local HuggingFace backends
+(transformers v5+) that reuse cached prefix computations across multiple
+generation calls.
 """
 
 from collections.abc import Iterable
-from functools import reduce
-from typing import Any
+from typing import Any, cast
 
 try:
     import torch
-    from transformers import PreTrainedModel
-    from transformers.cache_utils import DynamicCache
+    from transformers import PreTrainedModel, PreTrainedTokenizerBase
+    from transformers.cache_utils import CacheLayerMixin, DynamicCache
+    from transformers.generation.utils import GenerateDecoderOnlyOutput
     from transformers.tokenization_utils_base import BatchEncoding
 except ImportError as e:
     raise ImportError(
@@ -26,61 +27,110 @@
 LegacyCache = Any
 
 
-def legacy_cache_smash(a: LegacyCache, b: LegacyCache) -> LegacyCache:
-    """Concatenates two LegacyCache Ks and Vs along the time axis.
-
-    Args:
-        a: First legacy KV cache (tuple of per-layer (K, V) tensor pairs).
-        b: Second legacy KV cache to concatenate after `a`.
-
-    Returns:
-        New legacy cache with `b` appended to `a` along the sequence dimension.
-    """
-    legacy_merged = tuple(
-        (torch.cat([a[i][0], b[i][0]], dim=2), torch.cat([a[i][1], b[i][1]], dim=2))
-        for i in range(len(a))
+@torch.no_grad()
+def prefill_cache_v5(
+    model: PreTrainedModel,
+    tokenizer: PreTrainedTokenizerBase,
+    text: str,
+    device: torch.device,
+) -> tuple[dict, DynamicCache]:
+    """Prefills cache for transformers v5."""
+    toks = tokenizer(text, return_tensors="pt")
+    toks = {k: v.to(device) for k, v in toks.items()}
+
+    dc = DynamicCache()
+    out = model(
+        input_ids=toks["input_ids"],
+        attention_mask=toks["attention_mask"],
+        past_key_values=dc,
+        use_cache=True,
+    )
+    dc = out.past_key_values
+    dc.crop(-1)
+    return toks, dc  # v5 returns DynamicCache (not legacy tuple)
+
+
+def merge_dynamic_caches_v5(caches: Iterable[DynamicCache]) -> DynamicCache:
+    """Merge multiple v5 DynamicCache objects by concatenating KV states along the time axis."""
+    caches = list(caches)
+    if not caches:
+        raise ValueError("caches must be non-empty")
+
+    for c in caches:
+        if any(
+            getattr(layer, "is_sliding", False) for layer in getattr(c, "layers", [])
+        ):
+            raise ValueError(
+                "KV cache smashing does not currently support sliding-window "
+                "attention layers (e.g. Granite 4 models). Disable prefix "
+                "caching for this model."
+            )
+
+    merged = DynamicCache()
+
+    # reuse Cache.update() to append each segment's KV to the merged cache per layer.
+    # DynamicLayer.update(): self.keys = cat([self.keys, key_states], dim=-2).
+    for c in caches:
+        for layer_idx, layer in enumerate(c.layers):
+            if isinstance(layer, CacheLayerMixin):
+                if layer.keys is None or layer.values is None:
+                    continue
+                merged.update(layer.keys, layer.values, layer_idx=layer_idx)
+
+    return merged
+
+
+def merge_v5(
+    model: PreTrainedModel,
+    tokenizer: PreTrainedTokenizerBase,
+    strs: list[str],
+    device: torch.device,
+):
+    """Merges DynamicCache for transformers>=5.0.0."""
+    strs_toks, strs_dcs = [], []
+    for s in strs:
+        toks, dc = prefill_cache_v5(model, tokenizer, s, device)
+        strs_toks.append(toks)
+        strs_dcs.append(dc)
+
+    merged_toks = torch.cat([t["input_ids"] for t in strs_toks], dim=1)
+    merged_masks = torch.cat([t["attention_mask"] for t in strs_toks], dim=1)
+
+    merged_dc = merge_dynamic_caches_v5(strs_dcs)
+
+    return merged_toks, merged_masks, merged_dc
+
+
+if __name__ == "__main__":
+    from mellea.backends.huggingface import LocalHFBackend
+    from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
+
+    assert IBM_GRANITE_3_3_8B.hf_model_name is not None
+    backend = LocalHFBackend(model_id=IBM_GRANITE_3_3_8B.hf_model_name)
+    _model_raw, tokenizer, device = backend._model, backend._tokenizer, backend._device
+    model = cast(PreTrainedModel, _model_raw)
+
+    docs = [
+        "Nathan Fulton is expert in large language models, formal verification, and reinforcement learning. He holds a Ph.D. from Carnegie Mellon University's Computer Science Department and has worked at Amazon Web Services and IBM Research. He currently works at IBM Research - Cambridge.",
+        "IBM Research has a headquarters at 1101 Kitchawan Rd in Yorktown Heights and a Cambridge office at 314 Main Street in Cambridge, MA.",
+        "What is the address of Nathan's place of work?",
+    ]
+
+    merged_tokens, merged_masks, merged_cache = merge_v5(
+        model, tokenizer, docs, device=backend._device
+    )
+    input_ids = merged_tokens.to(device)
+    generate_out = cast(
+        GenerateDecoderOnlyOutput,
+        model.generate(  # type: ignore[operator]
+            input_ids=input_ids,
+            use_cache=True,
+            return_dict_in_generate=True,
+            past_key_values=merged_cache,
+            max_new_tokens=512,
+        ),
+    )
+    result = tokenizer.decode(
+        generate_out.sequences[0, input_ids.shape[1] :], skip_special_tokens=True
     )
-    return legacy_merged
-
-
-def merge_dynamic_caches(caches: Iterable[DynamicCache]) -> DynamicCache:
-    """Merges two DynamicCache Ks and Vs along the time axis.
-
-    Args:
-        caches: Iterable of `DynamicCache` objects to merge in order.
-
-    Returns:
-        A single `DynamicCache` with all caches concatenated along the sequence dimension.
-    """
-    legacies = [c.to_legacy_cache() for c in caches]  # type: ignore
-    assert len(legacies) >= 1
-    rv = DynamicCache.from_legacy_cache(reduce(legacy_cache_smash, legacies))  # type: ignore
-    return rv  # type: ignore
-
-
-def tokens_to_legacy_cache(
-    model: PreTrainedModel, device: str, tokens_or_cache: BatchEncoding | DynamicCache
-) -> Iterable[LegacyCache]:
-    """Prefills and returns Ks and Vs as a LegacyCache.
-
-    Args:
-        model: The HuggingFace model used for prefill.
-        device: Target device string (e.g. `"cuda"`, `"cpu"`).
-        tokens_or_cache: Either a `BatchEncoding` to prefill, or an existing
-            `DynamicCache` to convert directly.
-
-    Returns:
-        Legacy KV cache representation as a tuple of per-layer (K, V) tensor pairs.
-    """
-    if type(tokens_or_cache) is DynamicCache:
-        return tokens_or_cache.to_legacy_cache()  # type: ignore
-    else:
-        tokens = tokens_or_cache
-        dc = DynamicCache()
-        with torch.no_grad():
-            dc = model(
-                tokens["input_ids"].to(device),  # type: ignore
-                attention_mask=tokens["attention_mask"].to(device),  # type: ignore
-                past_key_values=dc,
-            ).past_key_values
-        return dc.to_legacy_cache()
+    print(result)
diff --git a/mellea/backends/openai.py b/mellea/backends/openai.py
index 81b2864db..af5a7a074 100644
--- a/mellea/backends/openai.py
+++ b/mellea/backends/openai.py
@@ -6,7 +6,7 @@
 import inspect
 import os
 from collections.abc import Coroutine, Sequence
-from typing import TYPE_CHECKING, Any, overload
+from typing import Any, overload
 
 import openai
 from openai.types.chat import ChatCompletion
@@ -64,9 +64,6 @@
     convert_tools_to_json,
 )
 
-if TYPE_CHECKING:
-    from transformers.tokenization_utils import PreTrainedTokenizer
-
 openai_ollama_batching_error = "json: cannot unmarshal array into Go struct field CompletionRequest.prompt of type string"
 
 format: None = None  # typing this variable in order to shadow the global format function and ensure mypy checks for errors
diff --git a/mellea/formatters/granite/base/util.py b/mellea/formatters/granite/base/util.py
index 60c8ca9fc..4ad210234 100644
--- a/mellea/formatters/granite/base/util.py
+++ b/mellea/formatters/granite/base/util.py
@@ -389,13 +389,11 @@ def generate_with_transformers(
         # Third Party
         import torch
 
-    # Input tokens must be passed to generate() as a positional argument, not a named
-    # argument.
     input_tokens = generate_input["input_tokens"]
     generate_input = generate_input.copy()
     del generate_input["input_tokens"]
 
-    generate_result = model.generate(input_tokens, **generate_input)  # type: ignore[operator]
+    generate_result = model.generate(inputs=input_tokens, **generate_input)  # type: ignore[operator]
 
     # Result is a a 2D tensor of shape (num responses, prompt + max generated tokens)
     # containing tokens, plus a tuple of <max generated tokens> tensors of shape
@@ -462,6 +460,7 @@ def generate_with_transformers(
                 all_logprobs[token_ix][response_tokens[token_ix]].item()
                 for token_ix in range(len(response_tokens))
             ]
+            assert isinstance(response_string, str)
             token_strings = [response_string[begin:end] for begin, end in token_offsets]
             token_bytes = [list(s.encode("utf-8")) for s in token_strings]
 
@@ -474,8 +473,8 @@ def generate_with_transformers(
                     torch.nan_to_num(all_logprobs, float("-inf")),
                     other_input["top_logprobs"],
                 )
-                top_k_token_strs = [
-                    [tokenizer.decode(t) for t in row_i] for row_i in top_k_indices
+                top_k_token_strs: list[list[str]] = [
+                    [str(tokenizer.decode(t)) for t in row_i] for row_i in top_k_indices
                 ]
                 top_logprobs = [
                     [
diff --git a/pyproject.toml b/pyproject.toml
index a4e74cf9b..76fcf334e 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -44,7 +44,7 @@ hf = [
     "datasets>=4.0.0",
     "llguidance",
     "peft>=0.18.1", # Native aLoRA support added in PEFT 0.18.0
-    "transformers>=4.53.2,<5",
+    "transformers>5.5.0,<6.0.0", # 5.5.0 broke granite 4.0
     "trl==0.19.1",
     "huggingface-hub>=0.33.4",
 ]
diff --git a/test/backends/test_document_rendering_unit.py b/test/backends/test_document_rendering_unit.py
index d1edaada8..87b49d815 100644
--- a/test/backends/test_document_rendering_unit.py
+++ b/test/backends/test_document_rendering_unit.py
@@ -253,15 +253,21 @@ async def test_watsonx_renders_documents_in_prompt():
 async def test_huggingface_renders_documents_in_prompt():
     """HuggingFace backend includes rendered documents in the chat template input."""
     import torch
+    from transformers import BatchEncoding
     from transformers.generation.utils import GenerateDecoderOnlyOutput
 
     mock_tokenizer = MagicMock()
     mock_model = MagicMock()
     mock_device = torch.device("cpu")
 
-    # apply_chat_template returns input_ids tensor
-    fake_input_ids = torch.tensor([[1, 2, 3, 4, 5]])
-    mock_tokenizer.apply_chat_template = MagicMock(return_value=fake_input_ids)
+    # apply_chat_template returns a BatchEncoding (v5 behavior with return_dict=True)
+    fake_batch = BatchEncoding(
+        {
+            "input_ids": torch.tensor([[1, 2, 3, 4, 5]]),
+            "attention_mask": torch.tensor([[1, 1, 1, 1, 1]]),
+        }
+    )
+    mock_tokenizer.apply_chat_template = MagicMock(return_value=fake_batch)
 
     # model.generate returns a real GenerateDecoderOnlyOutput (not a MagicMock)
     # to avoid AsyncIterator detection in send_to_queue
diff --git a/test/backends/test_huggingface.py b/test/backends/test_huggingface.py
index c24213732..c0d452ded 100644
--- a/test/backends/test_huggingface.py
+++ b/test/backends/test_huggingface.py
@@ -376,19 +376,33 @@ async def test_generate_with_lock(backend) -> None:
     memoized: dict[torch.Tensor, str] = dict()  # type: ignore[name-defined]
     gen_func = model.generate
 
-    def mock_func(input_ids, *args, **kwargs):
+    def _extract_inputs(inputs, args, kwargs):
+        # Callers in mellea/backends/huggingface.py use three different conventions
+        # (positional, inputs=, input_ids=); canonicalize to one tensor key.
+        if inputs is not None:
+            return inputs
+        if args:
+            return args[0]
+        return kwargs.get("input_ids", kwargs.get("inputs"))
+
+    def mock_func(inputs=None, *args, **kwargs):
         """Mocks the generate function. Must call `populate_mocked_dict` with each input that must be cached before using this."""
+        key_tensor = _extract_inputs(inputs, args, kwargs)
         for key, val in memoized.items():
-            if torch.equal(key, input_ids):
+            if torch.equal(key, key_tensor):
                 time.sleep(random.uniform(0.1, 0.5))  # Simulate a bit of work.
                 return val
         assert False, "did not get a cached response"
 
     # Safely create the dict.
-    def populate_mocked_dict(input_ids, *args, **kwargs):
+    def populate_mocked_dict(inputs=None, *args, **kwargs):
         """Generates the model output and adds to the memoized dict."""
-        output = gen_func(input_ids, *args, **kwargs)  # type: ignore
-        memoized[input_ids] = output
+        key_tensor = _extract_inputs(inputs, args, kwargs)
+        if inputs is not None:
+            output = gen_func(inputs, *args, **kwargs)  # type: ignore
+        else:
+            output = gen_func(*args, **kwargs)  # type: ignore
+        memoized[key_tensor] = output
         return output
 
     model.generate = Mock(side_effect=populate_mocked_dict)
diff --git a/test/backends/test_huggingface_filter_options.py b/test/backends/test_huggingface_filter_options.py
index 1ad98698f..9d05882e1 100644
--- a/test/backends/test_huggingface_filter_options.py
+++ b/test/backends/test_huggingface_filter_options.py
@@ -47,9 +47,10 @@ class _FakeTokenizer:
 # ---------------------------------------------------------------------------
 
 _EXPECTED_INTERNAL_VARS = {
-    # Named parameters wired by apply_chat_template / render_jinja_template
+    # Variables bound in the Jinja namespace by render_jinja_template (transformers v5+)
     "messages",
     "tools",
+    "documents",
     "add_generation_prompt",
     # Jinja environment globals from _compile_jinja_template
     "raise_exception",
diff --git a/uv.lock b/uv.lock
index 835a2b276..94e382427 100644
--- a/uv.lock
+++ b/uv.lock
@@ -23,7 +23,8 @@ name = "accelerate"
 version = "1.13.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "numpy" },
     { name = "packaging" },
     { name = "psutil" },
@@ -701,7 +702,7 @@ wheels = [
 
 [[package]]
 name = "click"
-version = "8.3.3"
+version = "8.4.1"
 source = { registry = "https://pypi.org/simple" }
 resolution-markers = [
     "python_full_version >= '3.15'",
@@ -711,9 +712,9 @@ resolution-markers = [
 dependencies = [
     { name = "colorama", marker = "(python_full_version < '3.12' and sys_platform == 'win32') or (python_full_version >= '3.14' and sys_platform == 'win32')" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/bb/63/f9e1ea081ce35720d8b92acde70daaedace594dc93b693c869e0d5910718/click-8.3.3.tar.gz", hash = "sha256:398329ad4837b2ff7cbe1dd166a4c0f8900c3ca3a218de04466f38f6497f18a2", size = 328061, upload-time = "2026-04-22T15:11:27.506Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/9b/98/518d8e5081007684232226f475082b30087d0f585e8457db087298259f49/click-8.4.1.tar.gz", hash = "sha256:918b5633eddf6b41c32d4f454bf0de810065c74e3f7dbf8ee5452f8be88d3e96", size = 353007, upload-time = "2026-05-22T04:08:37.769Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/ae/44/c1221527f6a71a01ec6fbad7fa78f1d50dfa02217385cf0fa3eec7087d59/click-8.3.3-py3-none-any.whl", hash = "sha256:a2bf429bb3033c89fa4936ffb35d5cb471e3719e1f3c8a7c3fff0b8314305613", size = 110502, upload-time = "2026-04-22T15:11:25.044Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/0d/67e5b4109ea4a837e80daa87c2c696711955e40449a97e8926672534def2/click-8.4.1-py3-none-any.whl", hash = "sha256:482be17c6991b8c19c5429a1e995d9b0efdbb63172824c41f99965dc0ade8ec2", size = 116639, upload-time = "2026-05-22T04:08:35.26Z" },
 ]
 
 [[package]]
@@ -1025,7 +1026,8 @@ dependencies = [
     { name = "filelock" },
     { name = "fsspec", extra = ["http"] },
     { name = "httpx" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "multiprocess" },
     { name = "numpy" },
     { name = "packaging" },
@@ -1047,7 +1049,7 @@ version = "9.14.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "click", version = "8.1.8", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
-    { name = "click", version = "8.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "lxml" },
     { name = "primp" },
 ]
@@ -1166,7 +1168,8 @@ dependencies = [
     { name = "docling-parse" },
     { name = "filetype" },
     { name = "httpx" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "lxml" },
     { name = "marko" },
     { name = "ocrmac", marker = "sys_platform == 'darwin'" },
@@ -1240,7 +1243,8 @@ source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "accelerate" },
     { name = "docling-core" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "jsonlines" },
     { name = "numpy" },
     { name = "pillow" },
@@ -1869,21 +1873,52 @@ wheels = [
 
 [[package]]
 name = "huggingface-hub"
-version = "0.36.2"
+version = "1.16.1"
 source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "python_full_version == '3.13.*'",
+    "python_full_version == '3.12.*'",
+]
 dependencies = [
-    { name = "filelock" },
-    { name = "fsspec" },
-    { name = "hf-xet", marker = "platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64'" },
-    { name = "packaging" },
-    { name = "pyyaml" },
-    { name = "requests" },
-    { name = "tqdm" },
-    { name = "typing-extensions" },
+    { name = "filelock", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "fsspec", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "hf-xet", marker = "(python_full_version >= '3.12' and python_full_version < '3.14' and platform_machine == 'AMD64') or (python_full_version >= '3.12' and python_full_version < '3.14' and platform_machine == 'aarch64') or (python_full_version >= '3.12' and python_full_version < '3.14' and platform_machine == 'amd64') or (python_full_version >= '3.12' and python_full_version < '3.14' and platform_machine == 'arm64') or (python_full_version >= '3.12' and python_full_version < '3.14' and platform_machine == 'x86_64')" },
+    { name = "httpx", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "packaging", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "pyyaml", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "tqdm", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "typer", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "typing-extensions", marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/7c/b7/8cb61d2eece5fb05a83271da168186721c450eb74e3c31f7ef3169fa475b/huggingface_hub-0.36.2.tar.gz", hash = "sha256:1934304d2fb224f8afa3b87007d58501acfda9215b334eed53072dd5e815ff7a", size = 649782, upload-time = "2026-02-06T09:24:13.098Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/48/0f/ed994dbade67a54407c28cab96ef845e0e6d25500be56aca6394f8bfc9dd/huggingface_hub-1.16.1.tar.gz", hash = "sha256:7f1dc4c5ec21aed69be630ad0c3378616be16f3de1a47b141c0e812965d9c832", size = 792534, upload-time = "2026-05-21T18:40:00.908Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/a8/af/48ac8483240de756d2438c380746e7130d1c6f75802ef22f3c6d49982787/huggingface_hub-0.36.2-py3-none-any.whl", hash = "sha256:48f0c8eac16145dfce371e9d2d7772854a4f591bcb56c9cf548accf531d54270", size = 566395, upload-time = "2026-02-06T09:24:11.133Z" },
+    { url = "https://files.pythonhosted.org/packages/49/79/621a7dbb80c70974f73a597275351ebe03ce5bc65cb5f8f4acb5859252bc/huggingface_hub-1.16.1-py3-none-any.whl", hash = "sha256:64340de934b9ce37857ef85a82de72f5629e8a270f9119eabb12bf495eb53c22", size = 668176, upload-time = "2026-05-21T18:39:58.596Z" },
+]
+
+[[package]]
+name = "huggingface-hub"
+version = "1.18.0"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "python_full_version >= '3.15'",
+    "python_full_version == '3.14.*'",
+    "python_full_version < '3.12'",
+]
+dependencies = [
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "filelock", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "fsspec", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "hf-xet", marker = "(python_full_version < '3.12' and platform_machine == 'AMD64') or (python_full_version >= '3.14' and platform_machine == 'AMD64') or (python_full_version < '3.12' and platform_machine == 'aarch64') or (python_full_version >= '3.14' and platform_machine == 'aarch64') or (python_full_version < '3.12' and platform_machine == 'amd64') or (python_full_version >= '3.14' and platform_machine == 'amd64') or (python_full_version < '3.12' and platform_machine == 'arm64') or (python_full_version >= '3.14' and platform_machine == 'arm64') or (python_full_version < '3.12' and platform_machine == 'x86_64') or (python_full_version >= '3.14' and platform_machine == 'x86_64')" },
+    { name = "httpx", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "packaging", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "pyyaml", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "tqdm", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "typer", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "typing-extensions", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/fb/d8/748ea0a47f0fa15227fe682f7a80826b4b7c096e4818044b8f56d6cb66d6/huggingface_hub-1.18.0.tar.gz", hash = "sha256:f0c5ecd1ef8c6a60f86f61ee278f2c1570ba9e279c9f54de9094210723b3613b", size = 812699, upload-time = "2026-06-05T09:26:33.401Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0b/03/40a05316cb6616e5b7efd7773656441ab04b4b022c2199e79bb4622a92a3/huggingface_hub-1.18.0-py3-none-any.whl", hash = "sha256:729be4a976fb706dcc02d176bcda8a3f32bdf21a294e8f4b3dda6fbcbc9c1ab1", size = 684411, upload-time = "2026-06-05T09:26:31.48Z" },
 ]
 
 [[package]]
@@ -2809,7 +2844,7 @@ resolution-markers = [
 ]
 dependencies = [
     { name = "aiohttp", version = "3.13.5", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
-    { name = "click", version = "8.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "fastuuid", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "httpx", marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "importlib-metadata", version = "8.7.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
@@ -3228,7 +3263,8 @@ all = [
     { name = "fastapi" },
     { name = "grpcio" },
     { name = "httpx" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "ibm-watsonx-ai" },
     { name = "langchain-core" },
     { name = "litellm", version = "1.83.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
@@ -3256,7 +3292,8 @@ backends = [
     { name = "accelerate" },
     { name = "boto3" },
     { name = "datasets" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "ibm-watsonx-ai" },
     { name = "litellm", version = "1.83.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "litellm", version = "1.83.14", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
@@ -3281,7 +3318,8 @@ granite-retriever = [
 hf = [
     { name = "accelerate" },
     { name = "datasets" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "llguidance" },
     { name = "peft" },
     { name = "transformers" },
@@ -3305,7 +3343,8 @@ server = [
     { name = "uvicorn" },
 ]
 switch = [
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
 ]
 telemetry = [
     { name = "cpex" },
@@ -3437,7 +3476,7 @@ requires-dist = [
     { name = "rouge-score" },
     { name = "sentence-transformers", marker = "extra == 'granite-retriever'" },
     { name = "smolagents", marker = "extra == 'tools'", specifier = ">=1.0.0" },
-    { name = "transformers", marker = "extra == 'hf'", specifier = ">=4.53.2,<5" },
+    { name = "transformers", marker = "extra == 'hf'", specifier = ">5.5.0,<6.0.0" },
     { name = "trl", marker = "extra == 'hf'", specifier = "==0.19.1" },
     { name = "typer", marker = "extra == 'cli'" },
     { name = "uvicorn", marker = "extra == 'server'" },
@@ -3896,7 +3935,7 @@ version = "3.9.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "click", version = "8.1.8", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
-    { name = "click", version = "8.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "joblib" },
     { name = "regex" },
     { name = "tqdm" },
@@ -4177,7 +4216,7 @@ version = "1.0.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "click", version = "8.1.8", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
-    { name = "click", version = "8.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "pillow" },
     { name = "pyobjc-framework-vision" },
 ]
@@ -4689,7 +4728,8 @@ version = "0.19.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "accelerate" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "numpy" },
     { name = "packaging" },
     { name = "psutil" },
@@ -6543,7 +6583,8 @@ name = "sentence-transformers"
 version = "5.4.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "numpy" },
     { name = "scikit-learn" },
     { name = "scipy" },
@@ -6701,19 +6742,20 @@ wheels = [
 
 [[package]]
 name = "smolagents"
-version = "1.24.0"
+version = "1.26.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "jinja2" },
     { name = "pillow" },
     { name = "python-dotenv" },
     { name = "requests" },
     { name = "rich" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/0a/1c/8d8e7f39f3586dc85a7f96681b08b7c1cfbd734b258a025f368632f2d666/smolagents-1.24.0.tar.gz", hash = "sha256:4d5028ffbd72aca85fb2e8daac52d03fb7d4290a9bf99dbc311dd65980f6b5de", size = 225246, upload-time = "2026-01-16T05:37:04.156Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/bd/85/3ca67bb57743434ac321821ea54cbdbf0d3dcc8d55f37199709e83141340/smolagents-1.26.0.tar.gz", hash = "sha256:4ec92313265f9cfbcabfc88e192b4bc4505f8475dc5f33dc872062fc567037bd", size = 239034, upload-time = "2026-05-29T05:08:44.732Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/76/7a/461ead649c088d30f7847ef1c70f245a9c188381ed09a264633831050497/smolagents-1.24.0-py3-none-any.whl", hash = "sha256:54853759d07a92a939f1ff59238b757fb1a153204c90b69324c8e9709025aa86", size = 155727, upload-time = "2026-01-16T05:37:02.903Z" },
+    { url = "https://files.pythonhosted.org/packages/89/8c/72bd5edc13288e3f27d4d9c1ef65adc0a68c950ee1fc6d3be270e1110f6c/smolagents-1.26.0-py3-none-any.whl", hash = "sha256:70e1cfb1576f782da93190ee31d9bb2659e5ca4bd84fda0c412e1f20498f28b6", size = 161466, upload-time = "2026-05-29T05:08:43.28Z" },
 ]
 
 [[package]]
@@ -6960,7 +7002,8 @@ name = "tokenizers"
 version = "0.22.2"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/73/6f/f80cfef4a312e1fb34baf7d85c72d4411afde10978d4657f8cdd811d3ccc/tokenizers-0.22.2.tar.gz", hash = "sha256:473b83b915e547aa366d1eee11806deaf419e17be16310ac0a14077f1e28f917", size = 372115, upload-time = "2026-01-05T10:45:15.988Z" }
 wheels = [
@@ -7167,23 +7210,23 @@ wheels = [
 
 [[package]]
 name = "transformers"
-version = "4.57.6"
+version = "5.10.2"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "filelock" },
-    { name = "huggingface-hub" },
+    { name = "huggingface-hub", version = "1.16.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
+    { name = "huggingface-hub", version = "1.18.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "numpy" },
     { name = "packaging" },
     { name = "pyyaml" },
     { name = "regex" },
-    { name = "requests" },
     { name = "safetensors" },
     { name = "tokenizers" },
     { name = "tqdm" },
+    { name = "typer" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/c4/35/67252acc1b929dc88b6602e8c4a982e64f31e733b804c14bc24b47da35e6/transformers-4.57.6.tar.gz", hash = "sha256:55e44126ece9dc0a291521b7e5492b572e6ef2766338a610b9ab5afbb70689d3", size = 10134912, upload-time = "2026-01-16T10:38:39.284Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/8d/38/d5f978bd5091019e89aef29b9a831f5cd70f2598963a3ead8b9570cab592/transformers-5.10.2.tar.gz", hash = "sha256:f9a44b9c8ca9ab1156b467f574d832ea066284299c2fd0ed84641ccb592751fc", size = 8799687, upload-time = "2026-06-04T18:43:49.119Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/03/b8/e484ef633af3887baeeb4b6ad12743363af7cce68ae51e938e00aaa0529d/transformers-4.57.6-py3-none-any.whl", hash = "sha256:4c9e9de11333ddfe5114bc872c9f370509198acf0b87a832a0ab9458e2bd0550", size = 11993498, upload-time = "2026-01-16T10:38:31.289Z" },
+    { url = "https://files.pythonhosted.org/packages/73/6f/e1564b0cc182afa05e219a8e09a8e770ffaab879b6b824b56c819bd221da/transformers-5.10.2-py3-none-any.whl", hash = "sha256:8a669db546f82c7c3618cb46ceb0f0afd89292bc70f319c058f8332ec63e268d", size = 11003830, upload-time = "2026-06-04T18:43:45.303Z" },
 ]
 
 [[package]]
@@ -7334,7 +7377,7 @@ source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "annotated-doc" },
     { name = "click", version = "8.1.8", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
-    { name = "click", version = "8.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "rich" },
     { name = "shellingham" },
 ]
@@ -7476,7 +7519,7 @@ version = "0.46.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "click", version = "8.1.8", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" },
-    { name = "click", version = "8.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
+    { name = "click", version = "8.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" },
     { name = "h11" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/1f/93/041fca8274050e40e6791f267d82e0e2e27dd165627bd640d3e0e378d877/uvicorn-0.46.0.tar.gz", hash = "sha256:fb9da0926999cc6cb22dc7cd71a94a632f078e6ae47ff683c5c420750fb7413d", size = 88758, upload-time = "2026-04-23T07:16:00.151Z" }