space: load Nemotron the normal way — transformers-native (no trust_remote_code), NO mamba_ssm/causal_conv1d. Those custom Triton CUDA kernels were the segfault (THCPModule_initExtension); native falls back to pure-torch Mamba on ZeroGPU.

requirements.txt

@@ -1,18 +1,16 @@
-# HF Space (Gradio SDK + ZeroGPU). Structure mirrors the org's working
-# NPCverse space: gradio.Server (gradio 6) + @app.api + app.launch(). That
-# needs transformers 5 (gradio 6 forbids huggingface-hub<1.0, which tf<5
-# pins); our checkpoint's auto_map + trust_remote_code makes tf5 use the
-# repo's modeling code regardless. bf16, no bitsandbytes.
-# gradio itself comes from the Space's sdk_version (6.18.0); don't double-pin.
+# HF Space (Gradio SDK + ZeroGPU), structured like the org's working spaces:
+# gradio.Server + @app.api + app.launch(). Load the model the NORMAL way —
+# transformers-native NemotronH, NO mamba_ssm / causal_conv1d. Those custom
+# Triton CUDA kernels are what segfaulted on ZeroGPU (THCPModule_initExtension);
+# working Nemotron ZeroGPU spaces never install them and let transformers fall
+# back to pure PyTorch for the Mamba ops. bf16, no bitsandbytes.
 spaces
 torch==2.10.0
 transformers>=5
 accelerate
 huggingface_hub>=1.2
+sentencepiece
 textual>=1.0
 rich>=13.0
 pyyaml>=6.0
 httpx>=0.27
-# Nemotron-H hard-imports mamba_ssm's triton kernels (torch 2.10 / cu12 / cp312).
-https://github.com/state-spaces/mamba/releases/download/v2.3.2.post1/mamba_ssm-2.3.2.post1+cu12torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
-https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.6.2.post1/causal_conv1d-1.6.2.post1+cu12torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl

space/app.py

@@ -57,15 +57,18 @@ WARDEN_ERR = "spaces package not present (not on a ZeroGPU Space)"
 
 if spaces is not None:
     try:
-        from mamba_ssm.ops.triton.layernorm_gated import rmsnorm_fn  # noqa: F401
         import torch
         from transformers import AutoModelForCausalLM, AutoTokenizer
 
-        tok = AutoTokenizer.from_pretrained(WARDEN_REPO, trust_remote_code=True)
+        # NO trust_remote_code: use transformers' NATIVE NemotronH, which falls
+        # back to pure-PyTorch Mamba ops when mamba_ssm isn't installed. The
+        # NVIDIA remote modeling code instead hard-requires mamba_ssm's Triton
+        # CUDA kernels, which segfault under ZeroGPU. This is how working
+        # Nemotron ZeroGPU spaces do it.
+        tok = AutoTokenizer.from_pretrained(WARDEN_REPO)
         model = AutoModelForCausalLM.from_pretrained(
             WARDEN_REPO,
             torch_dtype=torch.bfloat16,
-            trust_remote_code=True,
             low_cpu_mem_usage=True,
         )
         model.to("cuda")  # intercepted by ZeroGPU emulation; migrated per call