CompilationError Traceback (most recent call last)
in <cell line: 1>()
----> 1 trainer_stats = trainer.train()

25 frames
/usr/local/lib/python3.10/dist-packages/triton/compiler/compiler.py in make_ir(self, options, codegen_fns, context)
111
112 def make_ir(self, options, codegen_fns, context):
--> 113 return ast_to_ttir(self.fn, self, context=context, options=options, codegen_fns=codegen_fns)
114
115 def parse_options(self):

CompilationError: at 59:8:
# Do logit softcapping for Gemma 2: t * tanh(1/t * x)
if DO_SOFTCAPPING: logits = SOFTCAP * triton_tanh(logits / SOFTCAP)

c = tl.max(logits, 0)
logsumexp = c + tl.log(tl.sum(tl.exp(logits - c), 0))

if label_idx != -100:
    x = tl.load(logits_ptr + label_idx)
    # Go logit scaling for Cohere: t * x
    if DO_LOGIT_SCALING: x = LOGIT_SCALE * x
    # Do logit softcapping for Gemma 2: t * tanh(1/t * x)
    if DO_SOFTCAPPING:   x = SOFTCAP * triton_tanh(x / SOFTCAP)
    ^

AssertionError('initial value for x is of type fp16, but the then block redefines it as fp32')