fix(fireworks_ai): read fractional sizes like 1p5b in the size bucket fallback (#51451)

Fireworks writes decimal parameter counts with a 'p' (qwen3-0p6b, qwen2p5-1p5b). The fallback that picks a size bucket for models missing from the cost map only matched the digits right before the 'b', so 1p5b was read as 5B and billed in the 4.1B-16B bucket instead of the up-to-4B one.

Co-authored-by: Abelo9996 <Abelo9996@users.noreply.github.com>

litellm/llms/fireworks_ai/cost_calculator.py

@@ -38,11 +38,10 @@ def get_base_model_for_pricing(model_name: str) -> str:
         elif total_billion <= FIREWORKS_AI_176_B_MOE:
             return "fireworks-ai-56b-to-176b"
 
-    # Check for standard models in the form <number>b
-    re_params_match: Final = re.search(r"(\d+)b", model_name)
+    # Check for standard models in the form <number>b, where Fireworks writes 1.5b as 1p5b
+    re_params_match: Final = re.search(r"(\d+(?:p\d+)?)b", model_name)
     if re_params_match is not None:
-        params_match: Final = str(re_params_match.group(1))
-        params_billion: Final = float(params_match)
+        params_billion: Final = float(re_params_match.group(1).replace("p", "."))
 
         # Determine the category based on the number of parameters
         if params_billion <= FIREWORKS_AI_4_B:

tests/unit/llms/fireworks_ai/test_fireworks_ai_cost_calculator.py

@@ -365,6 +365,17 @@ def test_an_unknown_fireworks_model_still_falls_back_to_the_parameter_size_bucke
     assert prompt_cost > 0
 
 
+@pytest.mark.parametrize("size", ["0p6b", "1p5b", "3p8b"])
+def test_an_unknown_fireworks_model_with_a_fractional_size_falls_back_to_its_own_size_bucket(size: str):
+    usage: Final = _usage(prompt_tokens=100, cached_tokens=0, completion_tokens=10)
+
+    fractional_cost: Final = cost_per_token(model=f"accounts/acme/models/not-in-the-map-{size}", usage=usage)
+    whole_cost: Final = cost_per_token(model="accounts/acme/models/not-in-the-map-1b", usage=usage)
+
+    assert fractional_cost == whole_cost
+    assert fractional_cost[0] > 0
+
+
 _TIERED_INPUT_PATTERN: Final = re.compile(r"^input_cost_per_token_above_(\d+)k_tokens$")