giterator commented on code in PR #18336:
URL: https://github.com/apache/tvm/pull/18336#discussion_r2372517587
##########
python/tvm/relax/frontend/nn/llm/position_embedding.py:
##########
@@ -75,6 +75,51 @@ def rope_freq_gptj(s: tir.Var, d: tir.Var, d_range: int,
theta: float, dtype: st
return cos_freq, sin_freq, {freq_var: freq}
+def rope_freq_llama4( # pylint: disable=too-many-arguments,too-many-locals
+ s: tir.Var,
+ d: tir.Var,
+ d_range: int,
+ theta: float,
+ dtype: str,
+ factor: float,
+ low_freq_factor: float,
+ high_freq_factor: float,
+ original_max_position_embeddings: float,
+):
+ """Compute the inverse frequency of RoPE for llama3 RoPE scaling."""
+ orig_freq = tir.const(1, "float32") / tir.power(
+ theta, 2 * (d // 2) / tir.const(d_range, "float32")
+ )
+ orig_freq_var = tir.Var("orig_freq", "float32")
+
+ llama3_inv_scaling_factor = 1.0 / factor
+
+ if high_freq_factor == low_freq_factor:
+ wavelength = tir.const(2 * math.pi, "float32") / orig_freq_var
+ threshold_wavelen = tir.const(original_max_position_embeddings /
low_freq_factor, "float32")
+
+ scaled_freq = tir.if_then_else(
+ wavelength > threshold_wavelen, orig_freq_var / factor,
orig_freq_var
+ )
+ smoothed_freq = s * scaled_freq
+
+ else:
+ # Original smooth interpolation logic
+ inv_diff_freq_factor = 1.0 / (high_freq_factor - low_freq_factor)
+
+ llama3_alpha = original_max_position_embeddings / (2 * math.pi) *
inv_diff_freq_factor
+ llama3_beta = low_freq_factor * inv_diff_freq_factor
+ smooth = tir.max(0.0, tir.min(1.0, llama3_alpha * orig_freq_var -
llama3_beta))
+ smoothed_freq = s * (
+ (1.0 - smooth) * orig_freq_var * llama3_inv_scaling_factor +
smooth * orig_freq_var
+ )
Review Comment:
Fixed the variable naming for `rope_freq_llama4`
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
