Upload patch_granite_config.py with huggingface_hub

patch_granite_config.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+"""Adapt an APEX generate_config.sh output for Granite-4.0-H-Tiny (granitemoehybrid).
+
+Hybrid: 36 Mamba-2 layers (ssm_in/out/conv1d) + 4 attention layers (plain MHA) +
+MoE (64 routed + shared expert) on every layer. The stock generator emits
+ffn_*_exps/shexp, attn_q/k/v/output, and ssm_out — but MISSES the Mamba-2
+input/conv projections. We add ssm_in + ssm_conv1d at the mixer precision, and
+drop tensors Granite doesn't have. 1-D SSM state (ssm_a/d/dt/norm) stays F32.
+
+--ssm-type pins the Mamba-2 mixer projections (ssm_in/out/conv1d) higher (e.g.
+Q8_0) to test whether protecting the recurrence helps (the hand-roll variant).
+"""
+import argparse, re
+DROP = {"attn_gate","attn_qkv","shortconv.in_proj","shortconv.out_proj","ssm_alpha","ssm_beta"}
+ADD  = ["ssm_in","ssm_conv1d"]   # ssm_out already emitted by the generator
+SSM_ALL = ["ssm_in","ssm_out","ssm_conv1d"]
+
+def main():
+    ap=argparse.ArgumentParser(); ap.add_argument("infile"); ap.add_argument("outfile")
+    ap.add_argument("--ssm-type",default=None,help="override Mamba-2 mixer tensors (e.g. Q8_0)")
+    a=ap.parse_args()
+    out=[]
+    for l in (x.strip() for x in open(a.infile) if x.strip()):
+        m=re.match(r"blk\.\d+\.([\w.]+)=",l); tail=m.group(1) if m else ""
+        if tail in DROP: continue
+        # optionally bump ssm_out (already present) to the mixer override type
+        mo_out=re.match(r"(blk\.\d+\.ssm_out)=(\S+)",l)
+        if mo_out and a.ssm_type:
+            out.append(f"{mo_out.group(1)}={a.ssm_type}"); continue
+        out.append(l)
+        mo=re.match(r"blk\.(\d+)\.attn_output=(\S+)",l)
+        if mo:
+            n,t=mo.group(1),mo.group(2)
+            st=a.ssm_type or t
+            for name in ADD: out.append(f"blk.{n}.{name}={st}")
+    open(a.outfile,"w").write("\n".join(out)+"\n")
+    print(f"wrote {a.outfile} ({len(out)} lines)")
+if __name__=="__main__": main()