Speed up explainer: prebuilt Mamba kernels (transformers 5.8 + kernels hub), tighter answer budget

model_brick.py

@@ -248,7 +248,7 @@ if _should_load():
             prompt_len = inputs["input_ids"].shape[1]
             with torch.no_grad():
                 out = model.generate(
-                    **inputs, max_new_tokens=320, do_sample=False,
+                    **inputs, max_new_tokens=192, do_sample=False,
                     pad_token_id=tok.eos_token_id,
                 )
             return tok.decode(out[0][prompt_len:], skip_special_tokens=True).strip()

requirements.txt

@@ -2,12 +2,10 @@
 gradio==6.16.0            # gr.Server (FastAPI) + @app.api queue + ZeroGPU glue
 spaces                    # @spaces.GPU — ZeroGPU allocation on Hugging Face
 torch>=2.8.0              # ZeroGPU requirement (>=2.8)
-transformers>=4.51.0      # >=4.51 for Qwen3; also runs Nemotron (trust_remote_code)
+transformers>=5.8.0       # native NemotronH wired to hub kernels (>=5.8)
+kernels>=0.10.2           # pulls kernels-community/mamba-ssm + causal-conv1d
+                          # prebuilt for torch>=2.8 — replaces the naive Mamba
+                          # path (pip mamba-ssm wheels stop at torch 2.7 and
+                          # ZeroGPU build containers have no nvcc; never add
+                          # mamba-ssm/causal-conv1d here, the build will fail)
 accelerate                # device placement / efficient loading
-
-# Optional speed-up for the Nemotron Mamba-2 kernels. Left unpinned because
-# they compile against CUDA at build time and can fail; transformers falls back
-# to a pure-PyTorch path without them. If the model fails to boot, the clean
-# escape hatch is the env var:  FITCHECK_MODEL=Qwen/Qwen3-4B-Instruct-2507
-# mamba-ssm
-# causal-conv1d