🚀 Feature

Thunder needs to support torch.sym_max and torch.sym_min operations, which are used by TorchDynamo to represent max/min operations on symbolic values, particularly for sliding window attention patterns.

Current Behavior

Likely unsupported or falls back to eager evaluation.

Expected Behavior

Support symbolic max/min that preserves symbolic expressions:

• max(Sym(s0 - 1055), 0) → Sym(Max(0, s0 - 1055))
• Generate appropriate runtime checks

Example from torch.compile

sub: "Sym(s67 - 1056)" = l_kwargs_past_key_values_layers_0_cumulative_length - 1056
add_1: "Sym(s67 - 1055)" = sub + 1
sym_max: "Sym(Max(0, s67 - 1055))" = torch.sym_max(add_1, 0)

This pattern is used for computing KV-cache offsets with sliding windows.

Minimal Reproduction Case

import torch
import thunder

@thunder.jit
def compute_kv_offset(cumulative_length: int, sliding_window: int = 1056) -> int:
    """Compute offset for sliding window attention."""
    offset_raw = cumulative_length - sliding_window + 1
    # Should support: torch.sym_max for symbolic values
    kv_offset = max(offset_raw, 0)  # or torch.sym_max(offset_raw, 0)
    return kv_offset

assert compute_kv_offset(100) == 0    # max(100 - 1056 + 1, 0) = 0
assert compute_kv_offset(2000) == 945  # max(2000 - 1056 + 1, 0) = 945

print(compute_kv_offset._lc_cs.last_epilogue_traces[-1])
# def epilogue():
#   return 945

Use Case

Critical for HF transformer models with sliding window attention (e.g., Mistral, Llama with sliding window).

Success Criteria

• Support torch.sym_max(symbolic, constant)
• Support torch.sym_min(symbolic, constant)
• Support max(symbolic, constant) as alias
• Generated code produces correct runtime values
• Symbolic expression preserved for further optimization