tensorrt_llm/_torch/visual_gen/pipeline.py (3)

28-30: Replace en-dash characters with standard hyphens in docstring.

Static analysis flagged ambiguous "–" (en-dash) characters. Use "-" (hyphen-minus) for consistency and to avoid potential issues with ASCII-only tooling.

🔧 Proposed fix

-    * ``A-B``  – profile denoise steps A through B (same ``A-B`` format as LLM path)
-    * ``all``  – profile the full generation forward (denoise + VAE), skip warmup
-    * (unset)  – no profiler API calls; plain ``nsys profile`` captures everything
+    * ``A-B``  - profile denoise steps A through B (same ``A-B`` format as LLM path)
+    * ``all``  - profile the full generation forward (denoise + VAE), skip warmup
+    * (unset)  - no profiler API calls; plain ``nsys profile`` captures everything

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@tensorrt_llm/_torch/visual_gen/pipeline.py` around lines 28 - 30, Docstring
uses en-dash characters (–) in the bullet list which static analysis flagged;
replace them with standard hyphen-minus (-) characters. Edit the docstring in
tensorrt_llm._torch.visual_gen.pipeline (the multiline comment containing the
"A-B", "all" and "(unset)" bullets) and change each "–" to "-" so entries read
"* ``A-B``  - profile...", "* ``all``  - profile...", "* (unset)  - no
profiler...". Ensure only the punctuation characters are replaced and
formatting/markdown backticks remain unchanged.

---

41-46: Consider adding error handling for malformed input.

If the user provides a malformed value like "1-2-3" or "foo", the current code will raise an unhandled ValueError. This could be confusing for users. Consider adding validation with a clearer error message, or documenting the expected failure behavior.

🛡️ Optional: Add error handling

     # A-B format (same parser as LLM path)
     if "-" in val:
-        start, stop = val.split("-")
-        return int(start), int(stop)
+        parts = val.split("-")
+        if len(parts) != 2:
+            raise ValueError(
+                f"Invalid TLLM_PROFILE_START_STOP format: '{val}'. "
+                "Expected 'A-B', 'all', or a single step number."
+            )
+        try:
+            return int(parts[0]), int(parts[1])
+        except ValueError:
+            raise ValueError(
+                f"Invalid TLLM_PROFILE_START_STOP range: '{val}'. "
+                "Start and stop must be integers."
+            )
     # Single step
-    v = int(val)
-    return v, v
+    try:
+        v = int(val)
+        return v, v
+    except ValueError:
+        raise ValueError(
+            f"Invalid TLLM_PROFILE_START_STOP value: '{val}'. "
+            "Expected 'A-B', 'all', or a single step number."
+        )

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@tensorrt_llm/_torch/visual_gen/pipeline.py` around lines 41 - 46, The parsing
of the range string in the block that reads the variable val currently assumes
either "start-stop" or a single integer and will raise an unhandled ValueError
for inputs like "1-2-3" or "foo"; update this code to validate the format and
provide a clear error when malformed: if "-" in val, split and ensure there are
exactly two parts, attempt to cast both parts to int inside a try/except and
raise a ValueError with a descriptive message if casting fails or parts count !=
2; for the single-value branch, wrap int(val) in try/except and raise a clear
ValueError on failure. Ensure you keep the same return semantics (start, stop)
on success and reference the same variable names (val, start, stop) so callers
remain unchanged.

---

535-545: Minor style improvements suggested by static analysis.

1. Line 543: Loop variable name is unused - rename to _name
2. Lines 535, 545: Use tuple unpacking instead of concatenation

♻️ Proposed style fixes

         if extra_latents:
             extra_results = []
-            for name, (extra_latent, extra_decode_fn) in extra_latents.items():
+            for _name, (extra_latent, extra_decode_fn) in extra_latents.items():
                 extra_results.append(extra_decode_fn(extra_latent))
-            result = (primary_result,) + tuple(extra_results)
+            result = (primary_result, *extra_results)

And similarly for line 535:

             if extra_latents:
                 extra_results = [efn(elat) for _, (elat, efn) in extra_latents.items()]
-                result = (primary_result,) + tuple(extra_results)
+                result = (primary_result, *extra_results)

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@tensorrt_llm/_torch/visual_gen/pipeline.py` around lines 535 - 545, Rename
the unused loop variable name to _name in the extra_latents loop inside the
branch where self.rank == 0, and replace tuple concatenation usages with tuple
unpacking for clarity: when collecting extra_results (the block that calls
extra_decode_fn on extra_latent) change the loop to for _name, (extra_latent,
extra_decode_fn) in extra_latents.items(), and where the code currently builds
result using (primary_result,) + tuple(extra_results) (and the symmetric case
above), use (primary_result, *extra_results) instead; keep the else case
returning primary_result unchanged. Ensure references are to decode_fn,
extra_latents, extra_results, primary_result, and result.