feat(convert): Add decision.chat to support decision models that can also chat

Branch: GraniteSwitchDecision
AI-usage: full (Claude + Sonnet 5.5)
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gguf-py/gguf/constants.py

@@ -339,6 +339,7 @@ class Keys:
 
     class Decision:
         TYPE               = "{arch}.decision.type"
+        CHAT               = "{arch}.decision.chat"  # bool, the model also does text generation
         # note: single-use-case keys can be hard-coded in cpp code
         BLOCK_COUNT        = "{arch}.decision.block_count"
         ROUTING_BLOCK_COUNT = "{arch}.decision.routing_block_count"
@@ -6283,6 +6284,7 @@ class DecisionType:
     VLLM_D3 = "vllm-d3"  # same as pplx-decider, non-causal attention
     LFM2_D1 = "lfm2-d1"  # same as openjev, the labels depend on the question type
     LFM2_D1_OMNI = "lfm2-d1-omni"  # same head as laya on a bidirectional LFM2 trunk, other prompt layout
+    GRANITE_SWITCH = "granite-switch"  # same as openjev, run by the decision aLoRA of granite switch
 
 
 class VisionProjectorType:

gguf-py/gguf/gguf_writer.py

@@ -1386,6 +1386,9 @@ class GGUFWriter:
     def add_decision_type(self, value: str) -> None:
         self.add_string(Keys.Decision.TYPE.format(arch=self.arch), value)
 
+    def add_decision_chat(self, value: bool) -> None:
+        self.add_bool(Keys.Decision.CHAT.format(arch=self.arch), value)
+
     def add_decision_block_count(self, value: int) -> None:
         self.add_uint32(Keys.Decision.BLOCK_COUNT.format(arch=self.arch), value)