tensorrt_llm/models/quant_config_utils.py (1)

30-33: ⚡ Quick win

Consider defensive access for nested config keys.

If the HF config is malformed (missing group_0, weights, or input_activations), these direct accesses will raise a KeyError with a less helpful message than a validated ValueError. This is especially relevant for a public utility function that may receive varied external configs.

🛡️ Proposed fix to add defensive validation

     config_groups = hf_quant_config.get("config_groups")
     if config_groups is None:
         raise ValueError(f"config_groups is not set in {hf_quant_config}.")
 
-    weights_quant_config = config_groups["group_0"]["weights"]
-    inputs_quant_config = config_groups["group_0"]["input_activations"]
+    group_0 = config_groups.get("group_0")
+    if group_0 is None:
+        raise ValueError("config_groups must contain 'group_0'.")
+    weights_quant_config = group_0.get("weights")
+    inputs_quant_config = group_0.get("input_activations")
+    if weights_quant_config is None or inputs_quant_config is None:
+        raise ValueError("group_0 must contain 'weights' and 'input_activations'.")
     weights_quant_strategy = weights_quant_config["strategy"]
     inputs_quant_strategy = inputs_quant_config["strategy"]

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tensorrt_llm/models/quant_config_utils.py` around lines 30 - 33, Validate
presence of expected nested keys in config_groups before direct indexing: check
that "group_0" exists and that within it both "weights" and "input_activations"
exist; if any are missing, raise a clear ValueError describing which key is
missing rather than allowing a KeyError. Then safely read weights_quant_config =
config_groups["group_0"]["weights"] and inputs_quant_config =
config_groups["group_0"]["input_activations"], and proceed to extract
weights_quant_strategy and inputs_quant_strategy from those dicts. Reference the
variables config_groups, "group_0", weights_quant_config, inputs_quant_config,
weights_quant_strategy, and inputs_quant_strategy when implementing the
validation and error messages.

tests/unittest/models/test_quant_config_utils.py (1)

150-186: 💤 Low value

Consider extending parametrization to cover FP8 input strategy validation.

The current test cases cover unsupported weights strategies and NVFP4 input strategy mismatches, but don't exercise the FP8-specific input strategy validation (Lines 37-38 and 41-42 of the utility: channel requires "token", block requires "group").

🧪 Optional additional test cases

 `@pytest.mark.parametrize`(
     "weights,input_activations,error_match",
     [
         (
             {
                 "num_bits": 8,
                 "strategy": "tensor",
             },
             {
                 "num_bits": 8,
                 "strategy": "token",
             },
             "Unsupported weights_quant_strategy",
         ),
+        (
+            {
+                "num_bits": 8,
+                "strategy": "channel",
+            },
+            {
+                "num_bits": 8,
+                "strategy": "group",  # should be "token"
+            },
+            "Unsupported inputs_quant_strategy",
+        ),
+        (
+            {
+                "num_bits": 8,
+                "strategy": "block",
+            },
+            {
+                "num_bits": 8,
+                "strategy": "token",  # should be "group"
+                "group_size": 128,
+            },
+            "Unsupported inputs_quant_strategy",
+        ),
         (
             {
                 "num_bits": 4,

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/unittest/models/test_quant_config_utils.py` around lines 150 - 186, Add
parametrized cases to
test_update_quant_config_from_compressed_tensors_rejects_strategies that
exercise FP8-specific validation: call
update_quant_config_from_compressed_tensors (with QuantConfig() and
_compressed_tensors_config) using a weights config representing FP8 (e.g., type
"float" with 8 bits / FP8) and input_activations with the wrong strategies for
FP8 (one case using "channel" where FP8 requires "token", and one case using
"block" where FP8 requires "group"), and assert the appropriate ValueError match
messages for FP8 input strategy mismatches.