recipe: BPW footnote inputs (GGUF weight counts) + card scripts

recipe/pipeline/nex_aggregate.py

@@ -59,6 +59,15 @@ def sha_file(p):
     return h.hexdigest()
 
 
+def n_elements(p, prefix=None):
+    """Weight count of a GGUF (only tensors whose name starts with prefix, if given); None if unreadable."""
+    try:
+        reader = gguf.GGUFReader(p)
+    except (ValueError, OSError):
+        return None
+    return sum(int(t.n_elements) for t in reader.tensors if prefix is None or t.name.startswith(prefix))
+
+
 def tensors(p):
     """name -> (type, bytes); {} if the file cannot be read (e.g. still being written by a running quantize)."""
     out = {}
@@ -158,6 +167,8 @@ S["bf16"] = {
     "n_ctx": nctx,
     "scored_tokens": chunks * (nctx // 2 - 1) if chunks and nctx else None,
     "size_bytes": exists_size(bf_path),
+    "elements": n_elements(bf_path) if os.path.exists(bf_path) else None,
+    "bpw_logged": f(r"model size\s*=\s*[\d.]+ MiB \(([\d.]+) BPW\)", rd(f"{W}/logs/Q1_q106.log"), cast=str),
     "file": bf_fn,
     "readback": bf_line.split()[0] if bf_line else None,
     "arch": f(r"arch=(\S+)", bf_line, cast=str),
@@ -281,6 +292,8 @@ for im, st in TWINS:
         S["tiers"][im]["same_tensor_types_as_standard"] = TEN[st] == TEN[im]
     else:
         S["tiers"][im]["same_tensor_types_as_standard"] = None
+    # what that comparison covers: every tensor's name, type and byte size (tensors() above)
+    S["tiers"][im]["same_tensor_names_types_bytes_as_standard"] = S["tiers"][im]["same_tensor_types_as_standard"]
     sa, sb = S["tiers"][st]["size_bytes"], S["tiers"][im]["size_bytes"]
     S["tiers"][im]["file_size_delta_bytes"] = (sb - sa) if None not in (sa, sb) else None
     ha, hb = (sha_file(pa), sha_file(pb)) if os.path.exists(pa) and os.path.exists(pb) else (None, None)
@@ -292,6 +305,7 @@ S["aux"] = {mm_fn: exists_size(f"{W}/out/{mm_fn}")}
 S["mmproj"] = {
     "file": mm_fn,
     "size_bytes": S["aux"][mm_fn],
+    "elements": n_elements(f"{W}/out/{mm_fn}") if os.path.exists(f"{W}/out/{mm_fn}") else None,
     "readback": mm_line.split()[0] if mm_line else None,
     "arch": f(r"arch=(\S+)", mm_line, cast=str),
     "ftype": f(r"ftype=(\d+)", mm_line, cast=int),

recipe/pipeline/nex_render.py

@@ -102,7 +102,7 @@ PROTOCOL = (
     f"instruction), a unique nonce per request and `cache_prompt: false` (`cache_n = 0` asserted on "
     f"every timed request), 1 warm-up then the median of {REPS}. Decode numbers are the server's own "
     f"`predicted_per_second`. Box iced: no other model loaded.")
-TABLE_HEAD = ("| File | ftype | Size | BPW | KLD vs BF16 ↓² | Same top-1 ↑ | PPL (× BF16) | TG ROCm0 | TG Vulkan0 | PP ROCm0 |\n"
+TABLE_HEAD = ("| File | ftype | Size | BPW⁴ | KLD vs BF16 ↓² | Same top-1 ↑ | PPL (× BF16) | TG ROCm0 | TG Vulkan0 | PP ROCm0 |\n"
               "| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |")
 TG_NOTE = f"TG = decode tokens/s after the {PN_TXT}-token code prompt, no draft head. PP = prefill tokens/s on ROCm0."
 
@@ -121,7 +121,7 @@ def quality_blurb():
 
 def bf16_row():
     b = S.get("bf16") or {}
-    return (f"| *BF16 reference* | {g(b.get('ftype'), '{}')} | {gib(b.get('size_bytes'))}³ | 16 | 0 | 100 % | "
+    return (f"| *BF16 reference* | {g(b.get('ftype'), '{}')} | {gib(b.get('size_bytes'))}³ | {g(b.get('bpw_logged'), '{}')} | 0 | 100 % | "
             f"{pm(b.get('ppl_paired'), b.get('ppl_paired_err'))}¹ | — | — | — |")
 
 
@@ -131,7 +131,18 @@ def footnotes(where="below"):
             f"¹ The BF16 PPL shown is the paired base every \"×\" ratio is computed against (averaged over the same scored tokens "
             f"in the KL-divergence runs). The standalone BF16 run's own summary line reads {pm(b.get('ppl'), b.get('ppl_err'))}.\n"
             f"² Quality columns: see *Where the quality numbers come from* {where}.\n"
-            f"³ BF16 conversion of the checkpoint; not published.")
+            f"³ BF16 conversion of the checkpoint; not published.\n"
+            f"{bpw_note()}")
+
+
+def bpw_note():
+    b, m = S.get("bf16") or {}, S.get("mmproj") or {}
+    n, v, p = b.get("elements"), m.get("elements"), SRC.get("params")
+    s = f"⁴ BPW as printed by `llama-quantize`: bits per weight over the {g(n, '{:,}')} weights in each GGUF."
+    if None not in (n, v, p) and n + v == p:
+        s += (f" The {p:,}-parameter count above also includes the {v:,}-weight vision tower, which ships in the "
+              f"projector file.")
+    return s
 
 
 YAML = """---
@@ -643,8 +654,8 @@ def imat_card():
 Importance-matrix-calibrated ROCmFP4 quantizations of
 **[nex-agi/Nex-N2.5-mini](https://huggingface.co/nex-agi/Nex-N2.5-mini)** ({g(SRC.get('params'), '{:,}')} parameters,
 {intro_arch()}, text + image). Companion to the standard build **[{STD}](https://huggingface.co/{STD})** — the same
-three 4-bit tiers, same vision projector, same unpatched `d3ca537` server; the only difference is how each 4-bit
-block's scale was chosen. There is no MTP head on either repo.
+three 4-bit tiers, same vision projector, same unpatched `d3ca537` server; the only difference in the weights is how
+each 4-bit block's scale was chosen. There is no MTP head on either repo.
 
 ## What the imatrix changes
 

recipe/results_summary.json

@@ -45,6 +45,8 @@
     "n_ctx": 2048,
     "scored_tokens": 40920,
     "size_bytes": 69376636928,
+    "elements": 34660610688,
+    "bpw_logged": "16.01",
     "file": "Nex-N2.5-mini-BF16.gguf",
     "readback": "PASS",
     "arch": "qwen35moe",
@@ -680,6 +682,7 @@
         "rms_dp": 8.201
       },
       "same_tensor_types_as_standard": true,
+      "same_tensor_names_types_bytes_as_standard": true,
       "file_size_delta_bytes": 256,
       "differs_from_standard": true,
       "header_keys_only_in_imatrix": [
@@ -728,6 +731,7 @@
         "rms_dp": 7.783
       },
       "same_tensor_types_as_standard": true,
+      "same_tensor_names_types_bytes_as_standard": true,
       "file_size_delta_bytes": 256,
       "differs_from_standard": true,
       "header_keys_only_in_imatrix": [
@@ -776,6 +780,7 @@
         "rms_dp": 8.509
       },
       "same_tensor_types_as_standard": true,
+      "same_tensor_names_types_bytes_as_standard": true,
       "file_size_delta_bytes": 256,
       "differs_from_standard": true,
       "header_keys_only_in_imatrix": [
@@ -794,6 +799,7 @@
   "mmproj": {
     "file": "mmproj-Nex-N2.5-mini-BF16.gguf",
     "size_bytes": 902821920,
+    "elements": 446571248,
     "readback": "PASS",
     "arch": "clip",
     "ftype": 32,